# GGUF Capability and Evidence Catalog
This generated reference records exhaustive GGUF capability verdicts, closure statistics,
immutable artifact identities, and evidence gaps. It is intended for maintainers and
machine review; see [`build_from_gguf()`](api/build_from_gguf.md) for user instructions.
The [model catalog](model-catalog.md) covers HuggingFace model registrations. It does not
imply GGUF import or runtime support; those capability-specific claims live only here.
**Pinned source:** `ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` (2026-08-23T16:59:42Z).
| Census | Total | Closure |
|---|---:|---|
| Architectures | 148 | graph verdicts: {'deferred': 41, 'rejected': 2, 'supported': 105}; importable: 105; quantized import: {'rejected': 47, 'supported': 101}; runtime: {'deferred': 135, 'rejected': 2, 'supported': 11} |
| Active stored qtypes | 25 | 24 have an import route; 1 are explicitly deferred with no route |
| Serialized projector strings | 60 | {'graph-importable': 56, 'runtime-evidenced': 0} |
| Tokenizer pre identifiers | 87 | 56 semantic groups; route dispositions: {'deferred-compiled-semantics': 45, 'deferred-pinned-artifact-evidence': 4, 'deferred-pinned-artifact-mismatch': 7, 'validated-pinned-source': 31} |
`SUPPORTED` means the named capability is implemented and mechanically tested. `DEFERRED` means it is intentionally unavailable pending the stated work. `REJECTED` means the input or route is invalid by policy. Graph support controls export. The separate runtime verdict records pinned real-artifact validation, independent parity, and deterministic generation or stateful semantics; it never gates export of a faithfully represented graph and package contract. Tokenizer `copy` requires embedded ordered-vocabulary identity; `pinned-source` also binds the complete GGUF artifact, immutable Hub assets, reconstruction policy, semantic hashes, and representative token-ID vectors.
## Runtime evidence
The first low-cost architecture batch promotes GPT-2, GPT-NeoX/Pythia, MPT, OLMo, StarCoder, and StarCoder2 using 334,238,976 bytes of GGUF payload and 346,825,051
download bytes including tokenizer assets. Every route is explicit-float only. The
network-free selection, budget, exclusions, and fail-closed candidate reasons are recorded in
`testdata/evidence/gguf_low_cost_runtime_batch.json`.
| Evidence ID | GGUF identity | Config identity | Tokenizer identity | Runtime proof |
|---|---|---|---|---|
| `apertus-v1.1-1.5b-instruct-bf16-ort-genai-0.15.2` | `MrMeOrYou/Apertus-v1.1-1.5B-Instruct-GGUF@88c75ad49566d3c2157d03709bf772262c3241ed`
`Apertus-v1.1-1.5B-Instruct-BF16.gguf`
3,028,052,608 B
`f9ec154d0ec29dad1f6465b458b7f27bd25ad7b9a3899233ae98ca6d358501c2` | `swiss-ai/Apertus-v1.1-1.5B-Instruct@9e9d01154446a645d30f04174cf1515a38058be7` | `swiss-ai/Apertus-v1.1-1.5B-Instruct@9e9d01154446a645d30f04174cf1515a38058be7`
`chat_template.jinja` 5,250 B `4afab8361a4bd0c2994404e0b0851dbaad461a06e56bdfdb635aae3977473c19`, `special_tokens_map.json` 560 B `9f69883bd70fc5d8b55822799837a216d3ac4fb565e05256a0d4f9850404bbc5`, `tokenizer.json` 17,078,368 B `be12f4375d655cc740864e3a9041bcddd8477942f209d9e7f27f6c8767162638`, `tokenizer_config.json` 177,274 B `77b14a0664585c26065f07d7a4c852a4615c83348d9378e23def01957bbd3f57`
metadata `3097bd9f22efd32db9045c4705d978539dc8adeeae763324062a5e1a73fc24a5` | ONNX Runtime 1.29.0 `CPUExecutionProvider`; ort-genai 0.15.2; result=passed; full-logit; dynamic KV cache prefill, full-sequence replay, rollback, reorder, and 20 decode steps |
| `gpt2-q2-k-ort-genai-0.15.2` | `tensorblock/gpt2-GGUF@5b01870b15c4b2e43695d7f3f3bfb5b26106f23b`
`gpt2-Q2_K.gguf`
81,196,544 B
`4234545f917ec1df10dab4d926796a83422b68e9010d85a4c111b8b541f32892` | `openai-community/gpt2@607a30d783dfa663caf39e06633721c8d4cfcd7e` | `Xenova/gpt2@bf2c7f02e0b826c60d03af341171bde20893da66`
`special_tokens_map.json` 99 B `6f50ab5a5a509a1c309d6171f339b196a900dc9c99ad0408ff23bb615fdae7ad`, `tokenizer.json` 2,107,653 B `cda20b8ca044949aa07ac4078420c80d1a57139d5f9f33700e46fb2d891e7c66`, `tokenizer_config.json` 234 B `551e26ec611d8d0c8edc3ef72e518a38418cb71f40de1347dd486a595e1557d7`
metadata `b2417176025f8500d864004b0bf93b1403dc3c52238f6628f82fb0e3c498977e` | ONNX Runtime 1.29.0 `CPUExecutionProvider`; ort-genai 0.15.2; result=passed; full-logit; dynamic KV cache prefill, full-sequence replay, rollback, reorder, and 20 decode steps; Explicit-float correctness route; source GGUF blocks are dequantized. |
| `lfm2-350m-f16-ort-genai-0.15.2` | `LiquidAI/LFM2-350M-GGUF@8fdc9d526b7ed346b19257551b05816c7912ecc2`
`LFM2-350M-F16.gguf`
711,482,304 B
`379ffdcbf08147c0313f6f1ce7ff558a2bc935eda633f4b46c52347032419c42` | `LiquidAI/LFM2-350M@f37d3f5c8c5484bc01dad379a595cf4c68c4e70e` | `LiquidAI/LFM2-350M@73e3c253078a3b97c2e14b4c4665679f4d9b6d56`
`chat_template.jinja` 209 B `a805e50fed68938a076b07e2e602639611b50b1ced0e50f11eb92f1ba25be4dc`, `special_tokens_map.json` 434 B `742aefe2b7dec496e8caffdba03a75d0c1a9925d53bd3f3e0d388c96b591b6f4`, `tokenizer.json` 4,732,426 B `98cff83b4f6d7e9d8929bebc62b07e92cf1b3f99c80d16bafe8b84a75448f40b`, `tokenizer_config.json` 91,509 B `36f511115e9d8952cbc9d15d9a20dfa7ce7d1444940e5c1dc42a762020c99bf5`
metadata `e5626d605bb50bc53fdb0fbfcf374fb33dfbaa0cc698d9746ba1e9b0b7e6d07d` | ONNX Runtime 1.29.0 `CPUExecutionProvider`; ort-genai 0.15.2; result=passed; full-logit; hybrid convolution and KV state prefill, replay, rollback, reorder, and 20 decode steps |
| `pythia-70m-q2-k-ort-genai-0.15.2` | `mradermacher/pythia-70m-GGUF@52d6f045404c9f93418df2a0144d20c9de34316b`
`pythia-70m.Q2_K.gguf`
38,508,192 B
`8e331c8c8016bed8ff1863b78fafe51e86b2364f32c2d7f3e201687e081cf7f7` | `EleutherAI/pythia-70m@a39f36b100fe8a5377810d56c3f4789b9c53ac42` | `EleutherAI/pythia-70m@a39f36b100fe8a5377810d56c3f4789b9c53ac42`
`special_tokens_map.json` 99 B `6f50ab5a5a509a1c309d6171f339b196a900dc9c99ad0408ff23bb615fdae7ad`, `tokenizer.json` 2,113,710 B `c24618a1b3e6a38167beff1c72cffd126c3a66254347304b50547d12c5f25624`, `tokenizer_config.json` 396 B `70e38394e494931c6f773ba41e19460dd4436526b852207367f04341b4066d3f`
metadata `d5c722f646ff6462ac217da5e3514d1fa8b4a7b33aedced3daa8e7f00cc74f78` | ONNX Runtime 1.29.0 `CPUExecutionProvider`; ort-genai 0.15.2; result=passed; full-logit; dynamic KV cache prefill, full-sequence replay, rollback, reorder, and 20 decode steps; Explicit-float correctness route on the portable graph; source GGUF blocks are dequantized and the pinned tokenizer vocabulary is extended only with deterministic padding IDs present in the GGUF. |
| `qwen2.5-0.5b-instruct-q8-ort-genai-0.15.2` | `Qwen/Qwen2.5-0.5B-Instruct-GGUF@9217f5db79a29953eb74d5343926648285ec7e67`
`qwen2.5-0.5b-instruct-q8_0.gguf`
675,710,816 B
`ca59ca7f13d0e15a8cfa77bd17e65d24f6844b554a7b6c12e07a5f89ff76844e` | `Qwen/Qwen2.5-0.5B-Instruct@7ae557604adf67be50417f59c2c2f167def9a775` | `Qwen/Qwen2.5-0.5B-Instruct@a338b55dd21219a5f4da42bc11a9313d1a27d4cc`
`tokenizer.json` 7,031,645 B `c0382117ea329cdf097041132f6d735924b697924d6f6fc3945713e96ce87539`, `tokenizer_config.json` 7,308 B `5214600ee45ca2f887ce2eede8910378a0111ea99d657428bcbce94778e65a92`
metadata `8fc8ef848104e931f14ae03d9581699d54813a2ff952fb7caac0654e8aa27ee3` | ONNX Runtime 1.29.0 `CPUExecutionProvider`; ort-genai 0.15.2; result=passed; full-logit; dynamic KV cache prefill, replay, rollback, reorder, and 20 decode steps |
| `qwen3.5-moe-0.87b-q2-k-ort-genai-0.15.2` | `Flexan/kshitijthakkar-qwen3.5-moe-0.87B-d0.8B-GGUF@a9b8adbec2cc87479c772dac1944f313b4036c26`
`qwen3.5-moe-0.87B-d0.8B.Q2_K.gguf`
626,599,552 B
`e8a84df1a50ce65cf80c2b55bba8c6e80f913679fdf9e9439f2c3b52ef3145d5` | `kshitijthakkar/qwen3.5-moe-0.87B-d0.8B@e5b5b3d7c3cc5593196902fd3c23964e891a6ea6` | `kshitijthakkar/qwen3.5-moe-0.87B-d0.8B@e5b5b3d7c3cc5593196902fd3c23964e891a6ea6`
`chat_template.jinja` 7,755 B `273d8e0e683b885071fb17e08d71e5f2a5ddfb5309756181681de4f5a1822d80`, `tokenizer.json` 12,807,982 B `5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42`, `tokenizer_config.json` 16,709 B `49e2b6e395f959f077f1e992b338919c0d4a9732fc6e613995e06557f843500c`
metadata `45302b58b2086a666a874652d0e9e1d5b4b26e786ffbaf9362a4f902eba0b10d` | ONNX Runtime 1.29.0 `CPUExecutionProvider`; ort-genai 0.15.2; result=passed; full-logit; hybrid convolution, recurrent, and KV state prefill, replay, rollback, reorder, and 20 cache-threaded decode steps; Explicit-float correctness route only: source quantization is dequantized, dense MoE execution is opt-in, and the selected publisher marks this reduced checkpoint as low quality. Same-value full-logit comparison uses atol=0.35 because small backend differences can cross a routed-expert boundary; all greedy tokens match. |
| `smollm-135m-f16-onnxruntime-1.29.0` | `neopolita/smollm-135m-gguf@22cca988936eafe92908e7558907c3964e10bba7`
`ggml-model-f16.gguf`
270,885,504 B
`ec8c775c16944a7e4b5251f97b3f848500dcc3e701b0d492ce9055cea42138a2` | `HuggingFaceTB/SmolLM-135M@1d461723eec654e65efdc40cf49301c89c0c92f4` | `HuggingFaceTB/SmolLM-135M@1d461723eec654e65efdc40cf49301c89c0c92f4`
`special_tokens_map.json` 831 B `e786b595b9a23148bf1630df78d9037a048ea671e48bfd3549a1e3c233742bb3`, `tokenizer.json` 2,104,556 B `9ca9acddb6525a194ec8ac7a87f24fbba7232a9a15ffa1af0c1224fcd888e47c`, `tokenizer_config.json` 3,685 B `238ad6b60d48e471624ea70bc79e92f2611844d5016471fee8c167854bcb98e8`
metadata `46646ba36ecae43de6f9f649d217774b889e0fd405af92205319b882927493fc` | ONNX Runtime 1.29.0 `CPUExecutionProvider`; onnx-genai 1.29.0; result=passed; full-logit; dynamic KV cache prefill plus 20 cache-threaded decode steps |
| `smollm-135m-f16-ort-genai-0.15.2` | `neopolita/smollm-135m-gguf@22cca988936eafe92908e7558907c3964e10bba7`
`ggml-model-f16.gguf`
270,885,504 B
`ec8c775c16944a7e4b5251f97b3f848500dcc3e701b0d492ce9055cea42138a2` | `HuggingFaceTB/SmolLM-135M@1d461723eec654e65efdc40cf49301c89c0c92f4` | `HuggingFaceTB/SmolLM-135M@1d461723eec654e65efdc40cf49301c89c0c92f4`
`special_tokens_map.json` 831 B `e786b595b9a23148bf1630df78d9037a048ea671e48bfd3549a1e3c233742bb3`, `tokenizer.json` 2,104,556 B `9ca9acddb6525a194ec8ac7a87f24fbba7232a9a15ffa1af0c1224fcd888e47c`, `tokenizer_config.json` 3,685 B `238ad6b60d48e471624ea70bc79e92f2611844d5016471fee8c167854bcb98e8`
metadata `46646ba36ecae43de6f9f649d217774b889e0fd405af92205319b882927493fc` | ONNX Runtime 1.29.0 `CPUExecutionProvider`; ort-genai 0.15.2; result=passed; full-logit; ORT GenAI prefill plus 20 cache-threaded decode steps |
| `tiny-mpt-q2-k-ort-genai-0.15.2` | `tensorblock/tiny-mpt-random-remote-code-GGUF@c151eb3f349485ee8ae72841d7d90b9121d5baa2`
`tiny-mpt-random-remote-code-Q2_K.gguf`
8,734,304 B
`5627dcb0ff18f6f7200f83c0aed2056a6a7c86b5f2d865833e1b5f00b00e4daa` | `echarlaix/tiny-mpt-random-remote-code@85e64794e74a6fb2e71e7055c7e0188ccdd32905` | `echarlaix/tiny-mpt-random-remote-code@85e64794e74a6fb2e71e7055c7e0188ccdd32905`
`special_tokens_map.json` 99 B `6f50ab5a5a509a1c309d6171f339b196a900dc9c99ad0408ff23bb615fdae7ad`, `tokenizer.json` 2,113,738 B `3cf430678137c8491ca82fb7092ee49e44ad38857fffe1e4a4a5ed860139a5b8`, `tokenizer_config.json` 237 B `7671fbb5b3d610e6e11d4f5fc78d3a7716e8846112ac7e0f72124caedf887570`
metadata `f15522d34f33354bc96d36a73ad4619925240328342d84a156855251c20d43af` | ONNX Runtime 1.29.0 `CPUExecutionProvider`; ort-genai 0.15.2; result=passed; full-logit; dynamic KV cache prefill, full-sequence replay, rollback, reorder, and 20 decode steps; Explicit-float correctness route; source GGUF blocks are dequantized. |
| `tiny-olmo-q2-k-ort-genai-0.15.2` | `tensorblock/tiny-random-olmo-GGUF@d0ee9498d082d6dc3e730b9765e6e87a9bb5d995`
`tiny-random-olmo-Q2_K.gguf`
33,860,576 B
`be1c5a22ac0e75cd5874467ffd80bcd2c8500609d3bf7ccdb5b269373e4d6da4` | `hyper-accel/tiny-random-olmo@88675ef0caa5bd10ece810c0f2a79faa7724f536` | `hyper-accel/tiny-random-olmo@88675ef0caa5bd10ece810c0f2a79faa7724f536`
`special_tokens_map.json` 293 B `a6188c1e366f8ed715e60ff39c46a8c500fc33508e2affeb23e8c547c5853193`, `tokenizer.json` 2,115,417 B `a094266ac6c4982efba277bc251349a5a6d6ad37efb39a2a90f53d8be2a40a40`, `tokenizer_config.json` 5,372 B `78a839c7851f14f9fb30e664c2b46166dc0628f2900679e5ec160656f702edff`
metadata `25efe0090ffe5a6deb777743917fe572548a0f23357f74d8bd9e969cc911fd73` | ONNX Runtime 1.29.0 `CPUExecutionProvider`; ort-genai 0.15.2; result=passed; full-logit; dynamic KV cache prefill, full-sequence replay, rollback, reorder, and 20 decode steps; Explicit-float correctness route; source GGUF blocks are dequantized. |
| `tiny-starcoder-q2-k-ort-genai-0.15.2` | `RichardErkhov/bigcode_-_tiny_starcoder_py-gguf@fa6f9fdbc134d86d78a3ff9ce08ab14ba33d4718`
`tiny_starcoder_py.Q2_K.gguf`
103,899,456 B
`aa8c2170bb9172447baba14309916cfc0d901dbffaf10f1448f4f631e10c1f41` | `bigcode/tiny_starcoder_py@8547527bef0bc927268c1653cce6948c5c242dd1` | `bigcode/tiny_starcoder_py@8547527bef0bc927268c1653cce6948c5c242dd1`
`special_tokens_map.json` 532 B `0823292e24ea07b89317e9ede9d08da2a1b6c014290c06908a7ad04f1efd6719`, `tokenizer.json` 2,057,395 B `42b5a37ba11199f024f2b8873e1ecba98da33166e16f700bf7cb2304b0a5583f`, `tokenizer_config.json` 677 B `95684c52ad9a970dbbb17576ee2237cb62902c1eff6804c7c91a4d6219a4a6d7`
metadata `23379a715b3983ce0f1559645984431bc039c16eaf6e062b7e62bceac6fa64cd` | ONNX Runtime 1.29.0 `CPUExecutionProvider`; ort-genai 0.15.2; result=passed; full-logit; dynamic KV cache prefill, full-sequence replay, rollback, reorder, and 20 decode steps; Explicit-float correctness route; source GGUF blocks are dequantized. |
| `tiny-starcoder2-q2-k-ort-genai-0.15.2` | `tensorblock/tiny-random-starcoder2-GGUF@82c9eb61d1af6ea00dff834f1ff0620144b333e8`
`tiny-random-starcoder2-Q2_K.gguf`
68,039,904 B
`ab0a4b4e79c906520808db065a00a317c8b097e2176638b34088975eada6e0ed` | `hyper-accel/tiny-random-starcoder2@193576733055d2108dc9906d0da6e0806ad9be57` | `hyper-accel/tiny-random-starcoder2@193576733055d2108dc9906d0da6e0806ad9be57`
`special_tokens_map.json` 1,300 B `0fc9ac706a35e6337b19d484bc8f866b6a0ee7ad2509b03f7305dc0838c82d2c`, `tokenizer.json` 2,060,947 B `17fa145258b20c18287f1e3bd804e074cc13333f11984a2f5a2f11c5110437aa`, `tokenizer_config.json` 7,877 B `8149d8e6b21275ad2cc346885ad92c0e9b5aa3e28a78bb5d39b7febc7e52545d`
metadata `38608a4848dc8535113cff0312858853a0c8e29a5cdc145f258e4a9d84b35113` | ONNX Runtime 1.29.0 `CPUExecutionProvider`; ort-genai 0.15.2; result=passed; full-logit; dynamic KV cache prefill, full-sequence replay, rollback, reorder, and 20 decode steps; Explicit-float correctness route; source GGUF blocks are dequantized. The tiny source config declares BOS/EOS ID 50256 outside its 49152-token model vocabulary. |
| `hy3-iq1-m-mtp-runtime-unvalidated` | combined: `AngelSlim/Hy3-GGUF@31b453f4d9b647c74e4c4f5cba632eb512332c91/Hy3-IQ1_M-mtp.gguf` 91,756,066,272 B `c1fe984fef6f23fd9eb144e41ce6824e3941ed0a8b6b8f9f3e63da0c2a1320c2`
total 91,756,066,272 B; above 16 GiB
target-only discriminator: `Hy3-IQ1_M.gguf` 89,446,312,384 B `8c4195718dc24384a38de1c492f4bca7d21447789714d232b0a4d9cd0bb0c806` | `tencent/Hy3@a960ebc3da325ba167f069f76c41eb62c9280d22` `0c9daab42bff9cce1b6f058b10d7b730f76d583e583e28ad56e92b36373246f0` | `tencent/Hy3@a960ebc3da325ba167f069f76c41eb62c9280d22`; separately deferred; metadata `bff10dd3b99c0d4098b28fbeb1af46804184cefc2920895969b6f1ad2e548fa3` | status=runtime_unvalidated; graph/package hashes unclaimed; ORT 1.29.0 `CPUExecutionProvider`; ort-genai 0.15.2 source `ed5f4e87147731e5b07810f9f5c90103b3603cdf`; The smallest pinned public complete Hunyuan-V3 MTP artifact is 91,756,066,272 bytes, above the 17,179,869,184-byte bounded-artifact policy. The matching target-only discriminator is 89,446,312,384 bytes.
ORT GenAI 0.15.2 exposes no two-model draft/target binding, independent target/MTP cache orchestration, proposal rollback, or acceptance statistics.
The embedded tokenizer uses pre=hunyuan-dense, whose compiled llama.cpp behavior is not serialized in GGUF. This is separate from the real-artifact budget and downstream two-model runtime validation gaps.
Bounded headers prove that the -mtp file has 81 physical blocks with trailing block 80 while the target-only file has exactly 80 blocks. No payload values, standalone MTP forward, graph, or package is promoted as runtime evidence. |
| `qwen3.6-27b-mtp-runtime-unvalidated` | combined: `localweights/Qwen3.6-27B-MTP-Q4_K_M-Q8nextn-GGUF@bdc8bc1ca4d45e1152d54c4e5f1389bc7bf6e439/Qwen3.6-27B-MTP-Q4_K_M-Q8nextn.gguf` 16,998,719,584 B `6cdb1aabb6ee711938df4192938b51228df9274f01c8d7a5ff404915ad342a7c`
total 16,998,719,584 B; within 16 GiB
split: `ggml-org/Qwen3.6-27B-GGUF@8a7ee08e8b9bfb857107ecc25a5599d2f38b76f8/Qwen3.6-27B-Q4_K_M.gguf` 19,095,766,304 B `65b753ea835627f7b511143c6ceb976525c7f21f5df8c664bc0a9c23d1c49921`, `ggml-org/Qwen3.6-27B-GGUF@8a7ee08e8b9bfb857107ecc25a5599d2f38b76f8/mtp-Qwen3.6-27B-Q4_0.gguf` 1,680,270,560 B `3d593f9e2788d59bb30d6024706b1efd5219fea466b6397c46159e3540937173`
total 20,776,036,864 B; above 16 GiB | `Qwen/Qwen3.6-27B@6a9e13bd6fc8f0983b9b99948120bc37f49c13e9` `69db4eb7196bc8190813231b3018ca05d8c2e3abc7b1af19d55c157af44a9d9c` | `Qwen/Qwen3.6-27B@6a9e13bd6fc8f0983b9b99948120bc37f49c13e9`; separately deferred; metadata `8d1070c727a7e6a03726687aa5746acece0255060a2a61a905c4c2285d353b68` | status=runtime_unvalidated; graph/package hashes unclaimed; ORT 1.29.0 `CPUExecutionProvider`; ort-genai 0.15.2 source `ed5f4e87147731e5b07810f9f5c90103b3603cdf`; reduced direct-ORT coordinator {'accepted': 1, 'proposal_steps': 51, 'rejected': 50, 'rollbacks': 50}; ORT GenAI 0.15.2 constructs one Model per Generator and exposes no two-model draft/target binding, independent MTP-cache orchestration, proposal accept/reject rollback, or observable acceptance statistics. Its pinned release matrix places speculative decoding on the roadmap; Generator.rewind_to only rewinds the single target generator and is not an MTP acceptance loop.
The embedded tokenizer uses pre=qwen35, whose compiled llama.cpp behavior is not serialized in GGUF. This tokenizer limitation is separate from downstream two-model runtime validation.
The under-budget combined artifact proves immutable header, physical trailing block, and layout identity only. No payload values, production tensor mapping, quantized storage execution, standalone MTP forward, graph, or package is promoted as runtime evidence. A separate tiny direct-ORT coordinator exercises target/MTP cache threading, accept/reject rollback, replay, reorder, and target-only equality without upgrading the real-artifact status.
The 2026-08-28 qualification transfer fetched 2,167,418,880 of 16,998,719,584 bytes before measured throughput implied a multi-hour transfer ahead of any 32 GiB-host build/runtime attempt. The partial payload was deleted; this operational deferral does not restrict export support. |
Pinned DFlash/EAGLE3 source, artifact, tokenizer, graph, and package hashes live in `testdata/evidence/gguf_draft_runtime_evidence.json`: `dflash`: 32 target-only-equal greedy tokens; 18/75 accepted; 3 multi-token rounds; 5 rollbacks; source cosine=0.999879, relative-L2=0.005406; `eagle3`: 32 target-only-equal greedy tokens; 20/43 accepted; 7 multi-token rounds; 10 rollbacks; source tensors exact. Both use separate target/draft caches; higher-level runtime=`runtime_unvalidated`.
The committed real-pair evidence uses a test-only direct ORT coordinator that reads remapping metadata from the raw immutable draft GGUF and does not import `DraftPairRunner` or its transition helpers. Per-round DFlash and EAGLE3 traces bind proposal/remap tokens, proposal-logit hashes, accepted prefixes, correction tokens, target replay, target/draft cache states, final counters, and four execution-mutating discriminators. Target replay starts from an empty cache, and final speculative rounds never process past the requested token count. Beam reorder is reported unsupported for the batch-size-one reference coordinator rather than inferred.
Runtime support above is independent from tokenizer materialization support below.
### Fail-closed runtime evidence
| Evidence ID | Pinned candidate | Bounded result | Withheld runtime claims |
|---|---|---|---|
| `nemotron-h-moe-30b-iq2-xxs-runtime-blocker` | `bartowski/nvidia_Nemotron-3-Nano-30B-A3B-GGUF@1fc64d5b160654ec892df2708aa893b0e96e6491`
`nvidia_Nemotron-3-Nano-30B-A3B-IQ2_XXS.gguf`
18,010,755,296 B
`f3da710c046ce7cc6ff28a9b5f1a9153ac72e3f60603e51c7bb679d80716b58a` | config/tokenizer `nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16@bf77c3174f68ad409e1c2aa60daeb46e32d1c606`; GGUF tokenizer metadata `6089bcaf08b3fe0d49379ca7e85bd3c93e8705bac6130636425c159212971225`; result=blocked; 401 tensors / 31,577,940,288 parameters; ORT 1.29.0 / CPUExecutionProvider; ort-genai 0.15.2; The smallest model GGUF in the pinned 29-file repository revision is 18,010,755,296 bytes, above the 16 GiB bounded-artifact policy; explicit float16/float32 weights require 63,155,880,576/126,311,761,152 bytes.
The normal optimized CPU export has 37,142 nodes (40,167 before Mobius optimization) and 6,028 MatMul nodes because its truthful ReLU2 routed experts remain an ONNX loop. ORT 1.29 MoE/QMoE exposes ReLU but not ReLU2. QMoE has separate router_probs/router_weights for correction-biased selection with unbiased sigmoid mixing, while shared experts and optional latent projections can surround it; those are not fused-op blockers. The pinned 30B has no latent projection. QMoE does not support GGUF IQ2_XXS storage, and Mobius has no proven IQ2_XXS packer/kernel path.
ORT GenAI 0.15.2 discovers sparse/nonconsecutive KV and conv/recurrent slots, but derives recurrent_state names while this export uses ssm_state, does not beam-reorder recurrent state, and rejects nonzero recurrent-state rewind.
The GGUF tokenizer declares pre=pixtral, whose compiled llama.cpp behavior is not serialized; exact ORT tokenizer materialization is unavailable. | cached decode and deterministic generation, full-logit parity, package and report roundtrip, state replay, rollback, and reorder, tensor value closure |
| `glm-dsa-iq1-s-artifact-budget-blocker` | `unsloth/GLM-5.2-GGUF@abc55e72527792c6e77069c99b4cb7de16fa9f23`
`UD-IQ1_S/GLM-5.2-UD-IQ1_S-00001-of-00006.gguf` 9,423,744 B `46b6148389219ae45167cb8124fbb18ef7d432daf619b4faf9e06ea80d3f4777`
`UD-IQ1_S/GLM-5.2-UD-IQ1_S-00002-of-00006.gguf` 49,208,128,256 B `f2180207285e04fcaa5b8c53ba6e77ad5cc58666b6e7c6b04a5eded3fe8bef09`
`UD-IQ1_S/GLM-5.2-UD-IQ1_S-00003-of-00006.gguf` 49,684,417,024 B `b1c0c5a302cc8d5d9ea0bcd4467c01db72c26839f820f7e882079582ea0a8d2b`
`UD-IQ1_S/GLM-5.2-UD-IQ1_S-00004-of-00006.gguf` 49,396,052,864 B `a6a42da6975e29f89866dcde2956e9e50e6ea26635fb5063b74f3973f4f863b6`
`UD-IQ1_S/GLM-5.2-UD-IQ1_S-00005-of-00006.gguf` 49,246,275,936 B `a4a9851a50db533f21ef824e5d8038f04e6782e7d602d18e5fdd6643f68ccccb`
`UD-IQ1_S/GLM-5.2-UD-IQ1_S-00006-of-00006.gguf` 19,171,063,136 B `3b767f55df64e0432d52fcf1a14eb47a1ef3bbc91339e2ae220f38602237d7d7`
total 216,715,360,960 B | blocked; The smallest mainline GLM-5.2 GGUF is a 216,715,360,960-byte split set whose trunk also declares an unsupported routed DSA/MLA MTP block. | real-weight full-logit parity, runtime packaging, deterministic generation |
| `minimax-m2-iq1-s-artifact-budget-blocker` | `mradermacher/MiniMax-M2-i1-GGUF@2d4f9b1a86d32ce4dfc47db312c8d6fcae8d7b37`
`MiniMax-M2.i1-IQ1_S.gguf` 46,514,882,176 B `7bae986e3cd380c28c6177d612fce1d52373241a0dfa13a6fca25de79abf15fb`
total 46,514,882,176 B | blocked; The smallest immutable public MiniMax-M2 GGUF is 46,514,882,176 bytes; real-weight parity cannot enter the 16 GiB bounded evidence set. | real-weight full-logit parity, runtime packaging, deterministic generation |
| `mistral4-iq1-m-artifact-budget-blocker` | `unsloth/Mistral-Small-4-119B-2603-GGUF@bd93c721735aa32c035c0f19e738cb3371fd56ff`
`Mistral-Small-4-119B-2603-UD-IQ1_M.gguf` 32,306,941,632 B `40fcdee4869110938638c6b8bac253f442b196518d4623f4afdf4b885cd961c7`
total 32,306,941,632 B | blocked; The smallest immutable public Mistral4 GGUF is 32,306,941,632 bytes; real-weight parity cannot enter the 16 GiB bounded evidence set. | real-weight full-logit parity, runtime packaging, deterministic generation |
## Remaining route work
Every unresolved route is classified once from its authoritative registry. Exact reasons
remain machine-readable in `_route_census.py`; this table groups only shared next work.
| Category | Next batch | Routes | Dependencies |
|---|---|---|---|
| `artifact-unavailable` | `projector-artifact-discovery` | `projector:phi4` | complete immutable mmproj; component parity oracle |
| `artifact-unavailable` | `tokenizer-artifact-replacement` | `tokenizer:gpt-4o`, `tokenizer:granite-embed-multi-311m`, `tokenizer:jina-v1-en`, `tokenizer:llama4` | replacement complete artifact; matching official tokenizer source |
| `dependency-or-mobius-abi-blocked` | `architecture-abi-dependencies` | `architecture:qwen4exp` | dedicated graph topology; cache/state contract; synthetic parity |
| `dependency-or-mobius-abi-blocked` | `architecture-abi-dependencies` | `architecture:afmoe`, `architecture:arwkv7`, `architecture:bailingmoe3`, `architecture:chameleon`, `architecture:cogvlm`, `architecture:cohere2moe`, `architecture:deepseek2`, `architecture:deepseek2-ocr`, `architecture:deepseek32`, `architecture:deepseek4`, `architecture:gemma3n`, `architecture:gemma4-assistant`, `architecture:gpt-oss`, `architecture:granite_swa`, `architecture:graniteswitch`, `architecture:hunyuan_vl`, `architecture:laguna`, `architecture:llama4`, `architecture:mellum`, `architecture:mimo2`, `architecture:minimax-m3`, `architecture:mistral3`, `architecture:nanbeige`, `architecture:paddleocr`, `architecture:plamo3`, `architecture:pockettts`, `architecture:qwen3tts`, `architecture:qwen3vl`, `architecture:qwen3vlmoe`, `architecture:rwkv6`, `architecture:rwkv6qwen2`, `architecture:rwkv7`, `architecture:step35`, `architecture:wavtokenizer-dec` | exact metadata extraction; tensor closure; dedicated graph and parity |
| `dependency-or-mobius-abi-blocked` | `mtp-specialized-abi` | `mtp:bailingmoe3`, `mtp:cohere2moe`, `mtp:deepseek2`, `mtp:deepseek32`, `mtp:deepseek4`, `mtp:glm-dsa`, `mtp:mimo2`, `mtp:nemotron_h_moe`, `mtp:qwen35moe`, `mtp:qwen3next`, `mtp:step35` | specialized sidecar graph; routed/cache state ABI |
| `dependency-or-mobius-abi-blocked` | `projector-package-abi` | `projector:resampler` | Mobius dynamic processor-to-graph media shape contract |
| `dependency-or-mobius-abi-blocked` | `tokenizer-compiled-semantics` | `tokenizer:afmoe`, `tokenizer:bloom`, `tokenizer:chameleon`, `tokenizer:codeshell`, `tokenizer:command-r`, `tokenizer:dbrx`, `tokenizer:deepseek-coder`, `tokenizer:deepseek-llm`, `tokenizer:deepseek-v3`, `tokenizer:default`, `tokenizer:exaone`, `tokenizer:exaone-moe`, `tokenizer:falcon`, `tokenizer:gpt3-finnish`, `tokenizer:granite-docling`, `tokenizer:granite-embed-multi-97m`, `tokenizer:grok-2`, `tokenizer:hunyuan`, `tokenizer:hunyuan-dense`, `tokenizer:jais`, `tokenizer:jais-2`, `tokenizer:joyai-llm`, `tokenizer:kimi-k2`, `tokenizer:laguna`, `tokenizer:megrez`, `tokenizer:mellum2`, `tokenizer:minerva-7b`, `tokenizer:minicpm5`, `tokenizer:minimax-m2`, `tokenizer:mpt`, `tokenizer:olmo`, `tokenizer:poro-chat`, `tokenizer:refact`, `tokenizer:sarvam-moe`, `tokenizer:seed-coder`, `tokenizer:smaug-bpe`, `tokenizer:solar-open`, `tokenizer:stablelm2`, `tokenizer:starcoder`, `tokenizer:superbpe`, `tokenizer:tekken`, `tokenizer:trillion`, `tokenizer:viking`, `tokenizer:whitespace`, `tokenizer:youtu` | compiled pinned llama.cpp oracle; dispatch-equivalence fixture |
| `dependency-or-mobius-abi-blocked` | `tokenizer-compiled-semantics` | `tokenizer:bailingmoe`, `tokenizer:bailingmoe2`, `tokenizer:chatglm-bpe`, `tokenizer:cohere2moe`, `tokenizer:glm4`, `tokenizer:llada-moe`, `tokenizer:tiny_aya` | upstream tokenizer semantic parity; independently proven replacement reconstruction |
| `evidence-only` | `architecture-runtime-evidence` | `architecture:arcee`, `architecture:arctic`, `architecture:baichuan`, `architecture:bailingmoe`, `architecture:bert`, `architecture:bitnet`, `architecture:bloom`, `architecture:chatglm`, `architecture:codeshell`, `architecture:cohere2`, `architecture:command-r`, `architecture:dbrx`, `architecture:deci`, `architecture:deepseek`, `architecture:dots1`, `architecture:dream`, `architecture:ernie4_5`, `architecture:ernie4_5-moe`, `architecture:eurobert`, `architecture:exaone`, `architecture:falcon`, `architecture:falcon-h1`, `architecture:gemma`, `architecture:gemma-embedding`, `architecture:gemma2`, `architecture:gemma3`, `architecture:gemma4`, `architecture:glm-dsa`, `architecture:granite`, `architecture:granitehybrid`, `architecture:granitemoe`, `architecture:grok`, `architecture:grovemoe`, `architecture:hunyuan-dense`, `architecture:hunyuan-moe`, `architecture:hy_v3`, `architecture:internlm2`, `architecture:jais`, `architecture:jais2`, `architecture:jamba`, `architecture:jina-bert-v2`, `architecture:jina-bert-v3`, `architecture:kimi-k3`, `architecture:kimi-linear`, `architecture:lfm2moe`, `architecture:llada`, `architecture:llada-moe`, `architecture:llama-embed`, `architecture:maincoder`, `architecture:mamba`, `architecture:mamba2`, `architecture:minicpm`, `architecture:minicpm3`, `architecture:minimax-01`, `architecture:minimax-m2`, `architecture:mistral4`, `architecture:modern-bert`, `architecture:muse-glimmer`, `architecture:nemotron`, `architecture:nemotron_h`, `architecture:nemotron_h_moe`, `architecture:neo-bert`, `architecture:nomic-bert`, `architecture:nomic-bert-moe`, `architecture:olmo2`, `architecture:olmoe`, `architecture:openelm`, `architecture:orion`, `architecture:pangu-embedded`, `architecture:phi2`, `architecture:phi3`, `architecture:phimoe`, `architecture:plamo`, `architecture:plamo2`, `architecture:plm`, `architecture:qwen`, `architecture:qwen2moe`, `architecture:qwen2vl`, `architecture:qwen3`, `architecture:qwen35`, `architecture:qwen3moe`, `architecture:qwen3next`, `architecture:refact`, `architecture:rnd1`, `architecture:seed_oss`, `architecture:smallthinker`, `architecture:smollm3`, `architecture:stablelm`, `architecture:t5`, `architecture:t5encoder`, `architecture:talkie`, `architecture:xverse` | immutable representative GGUF; full-logit prefill and cached-decode parity; deterministic generation/state evidence |
| `evidence-only` | `mtp-runtime-evidence` | `mtp:hy_v3`, `mtp:qwen35` | target acceptance loop; cache-threaded draft/target parity |
| `evidence-only` | `projector-runtime-evidence` | `projector:adapter`, `projector:cogvlm`, `projector:deepseekocr`, `projector:deepseekocr2`, `projector:dots3note_a`, `projector:dots3note_v`, `projector:dots_ocr`, `projector:exaone4_5`, `projector:gemma3`, `projector:gemma3na`, `projector:gemma3nv`, `projector:gemma4a`, `projector:gemma4ua`, `projector:gemma4uv`, `projector:gemma4v`, `projector:glm4v`, `projector:glma`, `projector:granite4_vision`, `projector:granite_speech`, `projector:hunyuanvl`, `projector:idefics3`, `projector:internvl`, `projector:janus_pro`, `projector:kimik25`, `projector:kimivl`, `projector:ldp`, `projector:ldpv2`, `projector:lfm2`, `projector:lfm2a`, `projector:lightonocr`, `projector:llama4`, `projector:meralion`, `projector:mimo_audio`, `projector:mimovl`, `projector:minicpmv4_6`, `projector:minimax_m3`, `projector:mlp`, `projector:muse-glimmer`, `projector:musicflamingo`, `projector:nemotron_v2_vl`, `projector:paddleocr`, `projector:parakeet`, `projector:pixtral`, `projector:pockettts_spkenc`, `projector:qwen2.5o`, `projector:qwen2.5vl_merger`, `projector:qwen2a`, `projector:qwen2vl_merger`, `projector:qwen3a`, `projector:qwen3tts_spkenc`, `projector:qwen3vl_merger`, `projector:step3vl`, `projector:ultravox`, `projector:voxtral`, `projector:yasa2`, `projector:youtuvl` | paired text target; processor boundary; deterministic multimodal package execution |
| `intentionally-rejected` | `policy-rejections` | `architecture:bailingmoe2`, `architecture:clip`, `architecture:dots3note`, `architecture:exaone-moe`, `architecture:exaone4`, `architecture:glm4`, `architecture:glm4moe`, `architecture:gptj` | policy change plus independent correctness proof |
| `intentionally-rejected` | `policy-rejections` | `projector:pockettts_gen`, `projector:qwen3tts_gen` | sidecar role must become a valid projector contract |
| `intentionally-rejected` | `policy-rejections` | `mtp:bailingmoe2`, `mtp:dots3note`, `mtp:exaone-moe`, `mtp:exaone4`, `mtp:gemma4-assistant`, `mtp:glm4`, `mtp:glm4moe`, `mtp:graniteswitch`, `mtp:nemotron_h` | upstream executable ownership change |
Recent PR dependencies: #645 (merged) — authoritative tokenizer route inventory and compiled-semantics blockers; #651 (merged) — gemma4 tokenizer evidence merged into the authoritative route census; #652 (closed) — superseded by merged PR #656; #656 (merged) — qwen4exp route and complete sharded import merged, superseding PR #652; #675 (merged) — follow-up hardening for sharded GGUF evidence edge cases after PR #656; #672 (merged) — Qwen3.5 MoE runtime evidence is complete; exclude it from follow-up evidence batches; #674 (merged) — Nemotron-H MoE blocker evidence is complete; exclude it from follow-up evidence batches; #677 (merged) — the text architecture cohort is merged into the authoritative graph-route census; #678 (merged) — the generic projector cohort is merged into the authoritative projector census; #679 (merged) — Hunyuan V3 graph and MTP routes are merged into the authoritative route census; #680 (merged) — the MoE architecture cohort is merged into the authoritative graph-route census.
## Tokenizer evidence
| Evidence ID | GGUF identity | Official source | Exact tokenizer proof |
|---|---|---|---|
| `gemma4-e2b-iq2-native-tokenizer` | `unsloth/gemma-4-E2B-it-GGUF@0314792d7f1f7e229411f620751375812bb9faf2`
`gemma-4-E2B-it-UD-IQ2_M.gguf`
2,290,860,128 B
`3d95ada2a122c9c0b42803317239b64b262ac9226a307ff895b3d87eec0c2acd`
architecture `gemma4`; declared pre `absent`; effective pre `gemma4` | `google/gemma-4-E2B-it@3e22461f65e89153144f8adb70e3b8c2cc9845a7`
`config.json` 4,954 B `1b28f3d2c3100f6c594754b81107428bd7b822a7f48272ca681dae9d2ec38330`
`chat_template.jinja` 18,569 B `0a2c8073c878ab1da004bee933a998606537bbb62016310352c7285c3f01c5b5`, `tokenizer.json` 32,169,626 B `cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f`, `tokenizer_config.json` 3,082 B `9f4fec4b1dc6ecddf8f4a92e9caea5971c0e67d81309f3f9066a2bee8c362633` | validated identifiers `['gemma4']`
metadata `ba1926593b1ede5e53dd8a41a435cf2783a832cf68f05be90b019350ec60ab77`
tokens 262,144 `7905cebbe55e92782d7179b7341d19a4968fdb69e76de970b2d9ede21f1b880d`
merges 514,906 `930c8fe84d284f73233935b8dc5b1499a9810f59a8d64f580c1a5e9c123134de`
types `987bc200faf7bd20738013daab9cda6a005b2a2842d4f8be9449458663e0bdf9`; scores=262144
user-defined IDs `[('', 101), ('', 49), ('', 51), ('<\|"\|>', 52), ('<\|channel>', 100), ('<\|tool_call>', 48), ('<\|tool_response>', 50)]`; source added tokens=24 `d2197ea6f594928aa6479c7e27b5a1fd004710b8930784b0606d32966eebde94`
pipeline `{'decoder': '78bc5c572c20213e27daca8fd5993992c1007941901a5fcceb5325362e21db23', 'normalizer': 'bee32d134b0862217fbe58f6ab6ef6a6d89e0e0eb000a084f321aaca31c1929c', 'post_processor': 'a443939c6288561c027ce2908243332dc594e122007588752fd052d928755b35', 'pre_tokenizer': '75caeae5a427c06d64010230e6a10b4f7a08253b5b5be3ab5ca4c15cdbf791f5'}`; chat `241c50d86bdfe5e43307da87f559cd2416aacd67a8de46c15acc0105ef2200b7`
source IDs `0..262143`; no GGUF-only padding extension; rows=262,144
materialized `c440114eceefe1e87a662a77b472ccfa733f4c7950271009dba324d54b7c3583`
`Hello, world! 12345` → `[9259, 236764, 1902, 236888, 236743, 236770, 236778, 236800, 236812, 236810]`
` spaced text\n` → `[138, 169862, 138, 1005, 107]`
`\u4f60\u597d\uff0c\u4e16\u754c\uff01` → `[144626, 236900, 12811, 237354]`
`Caf\xe9 \u2014 \u03ba\u03cc\u03c3\u03bc\u03bf\u03c2 \U0001f680` → `[160319, 236859, 2192, 150665, 148148, 236743, 242015]`
`<\|channel>thought\nplan<\|tool_call>call:f{}` → `[100, 45518, 107, 15081, 101, 48, 6639, 236787, 236760, 16454, 49]`
`Hello, world! 12345` + specials → `[2, 9259, 236764, 1902, 236888, 236743, 236770, 236778, 236800, 236812, 236810]`
GGUF-native Gemma4 reconstruction; official copy rejected: its seven tool/channel tokens are all special instead of llama.cpp user-defined (with the tool-response EOG override), it omits GGUF BOS insertion and cleanup decoding, names rather than as EOS, and its chat template hash differs from the GGUF-native template
llama.cpp oracle `8d9af256337d1a501250f9bbf4c0859a654bddd6`: 480 cases `1def4cb26eb2c9b671921869822490944998bd135f5d8acaeaf30161f2b80bb1` |
| `gpt2-q4-tokenizer` | `QuantFactory/gpt2-GGUF@7eae6f079f0164bff66b86eea5159f7a368f9381`
`gpt2.Q4_0.gguf`
106,554,880 B
`d52ac7ed12e1f87cbc93473912f5c213d7c7d6f2a0112ea9d78533d0d7bd3632`
architecture `gpt2`; declared pre `gpt-2`; effective pre `gpt-2` | `openai-community/gpt2@607a30d783dfa663caf39e06633721c8d4cfcd7e`
`config.json` 665 B `0daed7749b4f02b8f76240d5444551d7b08712dab4d0adb8239c56ba823bb7b4`
`tokenizer.json` 1,355,256 B `8414cab924d8b9b33013f0d221c5862f365ee9be39c5c2bfae8a5a9e970478a6`, `tokenizer_config.json` 26 B `5e04eb606e3a1583530a42e36c2a6b6615c86f34fe77e44d9ddeb43ff940931f` | validated identifiers `['a.x-4.0', 'exaone4', 'gigachat', 'gpt-2', 'jina-de', 'jina-es', 'jina-v2-de', 'jina-v2-es', 'mellum', 'modern-bert', 'phi-2']`
metadata `b2417176025f8500d864004b0bf93b1403dc3c52238f6628f82fb0e3c498977e`
tokens 50,257 `2d4e96560e324abcfaeaac6d24016c22f354ef9da213432b451e9a860b21d508`
merges 50,000 `e707935c815087d8103fec742a07d7e8b50d1acf997ddcde53c73213db1141a4`
types `aeaf4bf6f00438b0ef9ee1edb6e58616d49b4735d520393af901b7ccf1e3a218`; scores=0
user-defined IDs `[]`; source added tokens=0 `None`
pipeline `{}`; chat `None`
source IDs `0..50256`; no GGUF-only padding extension; rows=50,257
materialized `8414cab924d8b9b33013f0d221c5862f365ee9be39c5c2bfae8a5a9e970478a6`
`Hello, world! 12345` → `[15496, 11, 995, 0, 17031, 2231]`
` spaced text\n` → `[220, 38980, 220, 2420, 198]`
`\u4f60\u597d\uff0c\u4e16\u754c\uff01` → `[19526, 254, 25001, 121, 171, 120, 234, 10310, 244, 45911, 234, 171, 120, 223]`
`Caf\xe9 \u2014 \u03ba\u03cc\u03c3\u03bc\u03bf\u03c2 \U0001f680` → `[34, 1878, 2634, 851, 7377, 118, 139, 234, 38392, 34703, 26517, 35558, 12520, 248, 222]`
`<\|endoftext\|>` → `[50256]` |
| `jina-v2-code-q8-tokenizer` | `ggml-org/jina-embeddings-v2-base-code-Q8_0-GGUF@05e79e9a6c8b99491e92ebb28d753268f8601e3c`
`jina-embeddings-v2-base-code-q8_0.gguf`
172,869,280 B
`3bd1722f09350209aa3ada93df55882666c58194bfbbbe81c30545d731cb4e7a`
architecture `jina-bert-v2`; declared pre `jina-v2-code`; effective pre `jina-v2-code` | `jinaai/jina-embeddings-v2-base-code@516f4baf13dec4ddddda8631e019b5737c8bc250`
`config.json` 1,216 B `e426aa684c7f9a95c5f020aa855faf93a24f065f5fad0c9e17b124670cabdea6`
`special_tokens_map.json` 280 B `06e405a36dfe4b9604f484f6a1e619af1a7f7d09e34a8555eb0b77b66318067f`, `tokenizer.json` 2,561,316 B `b01c78a902aa4facb2f47f95449f48e2f7bbfea5d2472ee2f6ce92323c6f86e5`, `tokenizer_config.json` 493 B `f477aeb15ff9f78d3c1ddf2361d2b0b8b20cf55220f839f29a37f3a18efddd89` | validated identifiers `['jina-v2-code']`
metadata `30161844cf4cd814a532f368f372a6ba0c7c2c7d86d9678f9816505122e889e5`
tokens 61,056 `5a8d8f6a0dad37e10cb27a75ece22d44e96259419bdfe8f4e334a676483d9f78`
merges 60,795 `3c0cca5349df26b361d55c0498b4a46fb52a92a05e185754cf709b8de160c5da`
types `821b8840b71bceb8d52cacc36f2cc574f0c2e73f8df002cc0936566e6ba20043`; scores=0
user-defined IDs `[]`; source added tokens=0 `None`
pipeline `{}`; chat `None`
source IDs `0..61055`; no GGUF-only padding extension; rows=61,056
materialized `b01c78a902aa4facb2f47f95449f48e2f7bbfea5d2472ee2f6ce92323c6f86e5`
`Hello, world! 12345` → `[10564, 16, 7550, 5, 53737]`
` spaced text\n` → `[225, 4113, 72, 225, 1460, 203]`
`\u4f60\u597d\uff0c\u4e16\u754c\uff01` → `[12552, 19692, 2397, 47406, 32039, 19513]`
`Caf\xe9 \u2014 \u03ba\u03cc\u03c3\u03bc\u03bf\u03c2 \U0001f680` → `[39, 1326, 2521, 25956, 31085, 20788, 10123, 59430, 14535, 30975]`
`def fibonacci(n: int) -> int:\n return n if n < 2 else fibonacci(n-1) + fibonacci(n-2)` → `[406, 23852, 267, 33404, 12, 82, 30, 577, 13, 1882, 577, 30, 287, 437, 321, 392, 321, 318, 491, 723, 23852, 267, 33404, 12, 82, 17, 21, 13, 464, 23852, 267, 33404, 12, 82, 17, 22, 13]`
`Hello, world! 12345` + specials → `[0, 10564, 16, 7550, 5, 53737, 2]` |
| `kanana2-1.3b-instruct-q8-tokenizer` | `dummy9996/kanana-2-1.3b-instruct-GGUF@6c998111f40f3ab7adf65620a6a752230d8c75f6`
`kanana-2-1.3b-instruct-Q8_0.gguf`
1,377,890,688 B
`0b63b6b68f0c1f0e667ad070808dfee7a03db06fc41c3fe23c9d794841c6f801`
architecture `qwen3`; declared pre `kanana2`; effective pre `kanana2` | `kakaocorp/kanana-2-1.3b-instruct@bf4786aa2a1908adce942d53976270132732f720`
`config.json` 2,019 B `fe14b20b4b616d62ca0682312c2fcd2b90d9a836d14a1ff6448db3f533fd15a1`
`chat_template.jinja` 10,725 B `b8ee6b31575eada17ebbe73d3f1ac65d3efde64f0a25ff922031dec7e1cae3e3`, `tokenizer.json` 10,057,457 B `1c4be9ecf77c926456fb82d4cf07ff1218a91907f3408f44895d2b01e0f2b5ab`, `tokenizer_config.json` 50,155 B `1cdee8fcd4f6209e07e6d9966c8a3ff2d738830d79475193e94e448e153ae2d5` | validated identifiers `['kanana2']`
metadata `94c64f4813926cc68c2357c49a4f264a4788422f5b663299de0d6da63c4546e4`
tokens 128,256 `ba8fcb1c6a9186257d3e12f93bd8b77a50378a8e081c524c9c97ff652d34c941`
merges 127,744 `04e73d514ad172c5f02929d7517a6a01a9b42630223e365f06b20d6628a8d1b6`
types `552780454e0de07e46ed452dba9c24001a780923e0ba1e4b4d0906cb02c2aeab`; scores=0
user-defined IDs `[]`; source added tokens=0 `None`
pipeline `{}`; chat `None`
source IDs `0..128255`; no GGUF-only padding extension; rows=128,256
materialized `1fba3871de7549016b48a7890d46403fe24c59cf39afc9df395ea02c199d1917`
`Hello, world! 12345` → `[17263, 11, 1666, 0, 220, 9654, 2995]`
` spaced text\n` → `[220, 49580, 220, 2620, 198]`
`\u4f60\u597d\uff0c\u4e16\u754c\uff01` → `[117006, 6936, 9428, 31645]`
`Caf\xe9 \u2014 \u03ba\u03cc\u03c3\u03bc\u03bf\u03c2 \U0001f680` → `[34, 3685, 1989, 3968, 77914, 70157, 23526, 29669, 98218, 106943, 222]`
`<\|im_start\|>user\n\uc548\ub155\ud558\uc138\uc694<\|im_end\|>` → `[128009, 3043, 198, 15191, 128010]`
`Hello, world! 12345` + specials → `[128000, 17263, 11, 1666, 0, 220, 9654, 2995]`
llama.cpp oracle `8d9af256337d1a501250f9bbf4c0859a654bddd6`: 444 cases `ca7875445f21a03eb9a480c6aa96251bf4a8951a6e284dc480ef32eaedb796f5` |
| `lfm2-350m-f16-tokenizer` | `LiquidAI/LFM2-350M-GGUF@8fdc9d526b7ed346b19257551b05816c7912ecc2`
`LFM2-350M-F16.gguf`
711,482,304 B
`379ffdcbf08147c0313f6f1ce7ff558a2bc935eda633f4b46c52347032419c42`
architecture `lfm2`; declared pre `lfm2`; effective pre `lfm2` | `LiquidAI/LFM2-350M@73e3c253078a3b97c2e14b4c4665679f4d9b6d56`
`config.json` 999 B `fd3b3fba4e50e7b9a22bd41cbab59e9b28e319b2de19668d7fd9777c8d1a9ba1`
`chat_template.jinja` 209 B `a805e50fed68938a076b07e2e602639611b50b1ced0e50f11eb92f1ba25be4dc`, `special_tokens_map.json` 434 B `742aefe2b7dec496e8caffdba03a75d0c1a9925d53bd3f3e0d388c96b591b6f4`, `tokenizer.json` 4,732,426 B `98cff83b4f6d7e9d8929bebc62b07e92cf1b3f99c80d16bafe8b84a75448f40b`, `tokenizer_config.json` 91,509 B `36f511115e9d8952cbc9d15d9a20dfa7ce7d1444940e5c1dc42a762020c99bf5` | validated identifiers `['falcon-h1', 'falcon3', 'jina-v5-nano', 'lfm2', 'llama-bpe', 'llama-v3', 'llama3', 'midm-2.0', 'pixtral']`
metadata `e5626d605bb50bc53fdb0fbfcf374fb33dfbaa0cc698d9746ba1e9b0b7e6d07d`
tokens 65,536 `c004fd0578dbfbff394335a7d5f95e78a8cdbbff6abc8c389ba2290637be58b6`
merges 63,683 `c70042d0b5969460432a218556522dedee908735a3e4cf70f27936353c5b3f65`
types `ffe1ea561257dc6e1f2c257b99b4913d63e9d6b896cf2f9da1a3d2cac316d4b4`; scores=0
user-defined IDs `[]`; source added tokens=0 `None`
pipeline `{}`; chat `None`
source IDs `0..65535`; no GGUF-only padding extension; rows=65,536
materialized `e7b7960966e2ed43a22b00431246cf820d5e2751bec58c44f0184cbe9b8d18c9`
`Hello, world! 12345` → `[36309, 521, 2031, 510, 730, 10293, 2637]`
` spaced text\n` → `[730, 56551, 730, 3304, 708]`
`\u4f60\u597d\uff0c\u4e16\u754c\uff01` → `[11754, 6400, 1198, 11370, 8668]`
`Caf\xe9 \u2014 \u03ba\u03cc\u03c3\u03bc\u03bf\u03c2 \U0001f680` → `[544, 2305, 860, 2180, 59955, 49122, 27443, 16883, 51332, 23805, 758, 732]`
`<\|startoftext\|><\|im_start\|>user\nHello<\|im_end\|>` → `[1, 6, 6423, 708, 36309, 7]` |
| `qwen2.5-0.5b-instruct-q8-tokenizer` | `Qwen/Qwen2.5-0.5B-Instruct-GGUF@9217f5db79a29953eb74d5343926648285ec7e67`
`qwen2.5-0.5b-instruct-q8_0.gguf`
675,710,816 B
`ca59ca7f13d0e15a8cfa77bd17e65d24f6844b554a7b6c12e07a5f89ff76844e`
architecture `qwen2`; declared pre `qwen2`; effective pre `qwen2` | `Qwen/Qwen2.5-0.5B-Instruct@a338b55dd21219a5f4da42bc11a9313d1a27d4cc`
`config.json` 659 B `18e18afcaccafade98daf13a54092927904649e1dd4eba8299ab717d5d94ff45`
`tokenizer.json` 7,031,645 B `c0382117ea329cdf097041132f6d735924b697924d6f6fc3945713e96ce87539`, `tokenizer_config.json` 7,308 B `5214600ee45ca2f887ce2eede8910378a0111ea99d657428bcbce94778e65a92` | validated identifiers `['deepseek-r1-qwen', 'f2llmv2', 'kormo', 'qwen2']`
metadata `8fc8ef848104e931f14ae03d9581699d54813a2ff952fb7caac0654e8aa27ee3`
tokens 151,936 `e2fadeac783c911f535d21f858f43127672a1d261af510d3f895e34bd2f6fb10`
merges 151,387 `24fa2ae2a398e50784a1fff678482094af4f63e6783d35686726abacda8dc371`
types `17ccfa7767a8721474dc0fd21ca1308fdfd04e0f64036efbfa97f3e16e5f18f1`; scores=0
user-defined IDs `[]`; source added tokens=0 `None`
pipeline `{}`; chat `None`
source IDs `0..151935`; no GGUF-only padding extension; rows=151,936
materialized `be55f66f0643df9d3c1b5dc55ae552b0e334f219a3a5f8338e6864f8eb3a8ac5`
`Hello, world! 12345` → `[9707, 11, 1879, 0, 220, 16, 17, 18, 19, 20]`
` spaced text\n` → `[220, 63828, 220, 1467, 198]`
`\u4f60\u597d\uff0c\u4e16\u754c\uff01` → `[108386, 3837, 99489, 6313]`
`Caf\xe9 \u2014 \u03ba\u03cc\u03c3\u03bc\u03bf\u03c2 \U0001f680` → `[34, 2577, 963, 1959, 71638, 75195, 43928, 43123, 27554, 45642, 11162, 248, 222]`
`<\|im_start\|>user\nHello<\|im_end\|>\n` → `[151644, 872, 198, 9707, 151645, 198]` |
| `qwen3.5-0.8b-q4-tokenizer` | `ggml-org/Qwen3.5-0.8B-GGUF@8fea620810c4afa23dd6443f999a48574c1611a3`
`Qwen3.5-0.8B-Q4_0.gguf`
563,036,064 B
`57d1997790d1744fba5b40a7317df71ea5e2acee28c47e78f0cce39c0703f8cf`
architecture `qwen35`; declared pre `qwen35`; effective pre `qwen35` | `Qwen/Qwen3.5-0.8B@2fc06364715b967f1860aea9cf38778875588b17`
`config.json` 2,907 B `b90b86f35c8e6925ef74ee04d0e758f0a845c83a42089ad82bbaa948de9b4204`
`chat_template.jinja` 7,755 B `273d8e0e683b885071fb17e08d71e5f2a5ddfb5309756181681de4f5a1822d80`, `tokenizer.json` 12,807,982 B `5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42`, `tokenizer_config.json` 16,709 B `49e2b6e395f959f077f1e992b338919c0d4a9732fc6e613995e06557f843500c` | validated identifiers `['qwen35']`
metadata `45302b58b2086a666a874652d0e9e1d5b4b26e786ffbaf9362a4f902eba0b10d`
tokens 248,320 `5ee0f927bcaa4b9fe85c244776ae9487468e427f83e053fc81f2a186f14936a3`
merges 247,587 `7e299304d9ad9dc312acdbcb1f6ccf0dce1256bf1aa986d651f13814dfd27e7b`
types `f6fdca1063d1ae1cc77ba1f5087d259f044c2634e64b65e31bc844ec00e9acab`; scores=0
user-defined IDs `[]`; source added tokens=0 `None`
pipeline `{}`; chat `None`
source IDs `0..248076`; unused `[PAD{id}]` IDs `248077..248319`; rows=248,320
materialized `a78b900eb4cd335bba249158066db523ce221f744e2b6144692bb81673d551af`
`Hello, world! 12345` → `[9419, 11, 1814, 0, 220, 16, 17, 18, 19, 20]`
` spaced text\n` → `[220, 61674, 220, 1414, 198]`
`\u4f60\u597d\uff0c\u4e16\u754c\uff01` → `[109266, 3709, 96748, 6115]`
`Caf\xe9 \u2014 \u03ba\u03cc\u03c3\u03bc\u03bf\u03c2 \U0001f680` → `[34, 2492, 933, 1892, 166265, 203260, 10838, 248, 222]`
`<\|im_start\|>user\nHello<\|im_end\|>\n<\|im_start\|>assistant\n` → `[248045, 846, 198, 9419, 248046, 198, 248045, 74455, 198]`
`<\|audio_start\|><\|audio_pad\|><\|audio_end\|>` → `[248070, 248076, 248071]` |
| `roberta-bpe-q2-tokenizer` | `mradermacher/quora-roberta-base-GGUF@7a6d5816bb01c2d917978fb36825d9fec3ce4ff4`
`quora-roberta-base.Q2_K.gguf`
67,888,768 B
`1d31ba38f70d6f1456cfbd10c48dc6100a11a9b90558e110a9fb4d940b77cb49`
architecture `bert`; declared pre `roberta-bpe`; effective pre `roberta-bpe` | `sentence-transformers/stsb-roberta-base@32d471df2968a46d1fe447d66a9275e8e63fcf12`
`config.json` 672 B `05fac50b3f0e2782f88ba1349ede146230005edef945fb336eeb6f9a8d815940`
`special_tokens_map.json` 239 B `378eb3bf733eb16e65792d7e3fda5b8a4631387ca04d2015199c4d4f22ae554d`, `tokenizer.json` 1,355,881 B `33465117406b9007673e8ba283f7f1383d9b5094df947481af60eec94ed7d7bd`, `tokenizer_config.json` 1,172 B `5992009790ef0a4ba5910d1e0dc04c4e0601d416131501080c694231548bf666` | validated identifiers `['roberta-bpe']`
metadata `9bc381b15c316f8ced2658ec079c0b2d5ea6c6dcddb615f2a6966bbb717bde74`
tokens 50,265 `db935e2c7440742d76167108001403b9c51be6de99a5f677d70be0771e446cf5`
merges 50,000 `e707935c815087d8103fec742a07d7e8b50d1acf997ddcde53c73213db1141a4`
types `70211118fea54968b4980f036554fd37269fc3c205dddbfc2b21004f570df7c3`; scores=0
user-defined IDs `[]`; source added tokens=0 `None`
pipeline `{}`; chat `None`
source IDs `0..50264`; no GGUF-only padding extension; rows=50,265
materialized `33465117406b9007673e8ba283f7f1383d9b5094df947481af60eec94ed7d7bd`
`Hello, world! 12345` → `[31414, 6, 232, 328, 17072, 1898]`
` spaced text\n` → `[1437, 42926, 1437, 2788, 50118]`
`\u4f60\u597d\uff0c\u4e16\u754c\uff01` → `[47856, 21402, 48975, 10809, 43251, 4394, 14285, 46015, 25448, 49127, 14285, 43251, 4394, 10172]`
`Caf\xe9 \u2014 \u03ba\u03cc\u03c3\u03bc\u03bf\u03c2 \U0001f680` → `[347, 2001, 1140, 93, 43662, 3070, 45704, 14285, 47721, 47049, 46122, 47756, 8103, 15113, 7471]`
`The quick brown fox jumps over the lazy dog.` → `[133, 2119, 6219, 23602, 13855, 81, 5, 22414, 2335, 4]`
`Hello, world! 12345` + specials → `[0, 31414, 6, 232, 328, 17072, 1898, 2]` |
| `smollm-135m-f16-tokenizer` | `neopolita/smollm-135m-gguf@22cca988936eafe92908e7558907c3964e10bba7`
`ggml-model-f16.gguf`
270,885,504 B
`ec8c775c16944a7e4b5251f97b3f848500dcc3e701b0d492ce9055cea42138a2`
architecture `llama`; declared pre `smollm`; effective pre `smollm` | `HuggingFaceTB/SmolLM-135M@1d461723eec654e65efdc40cf49301c89c0c92f4`
`config.json` 724 B `a1fe6f43e20f7a6c6dbc6380222af9526b5cef262446391a281c038249e3e3b7`
`special_tokens_map.json` 831 B `e786b595b9a23148bf1630df78d9037a048ea671e48bfd3549a1e3c233742bb3`, `tokenizer.json` 2,104,556 B `9ca9acddb6525a194ec8ac7a87f24fbba7232a9a15ffa1af0c1224fcd888e47c`, `tokenizer_config.json` 3,685 B `238ad6b60d48e471624ea70bc79e92f2611844d5016471fee8c167854bcb98e8` | validated identifiers `['smollm']`
metadata `46646ba36ecae43de6f9f649d217774b889e0fd405af92205319b882927493fc`
tokens 49,152 `ecc2f33f7cdf683196646ea97b005f82398e5ddbb0e143fbe95a402277eb1788`
merges 48,900 `3d6f4016bc9b70ea16f0f01b1dadb4504ad99c5eaa8584b81997dc65168e136b`
types `3a92d63c9763834e17f2d93490d5a9643fa07057f2799168d67d7812d08e31aa`; scores=0
user-defined IDs `[]`; source added tokens=0 `None`
pipeline `{}`; chat `None`
source IDs `0..49151`; no GGUF-only padding extension; rows=49,152
materialized `9ca9acddb6525a194ec8ac7a87f24fbba7232a9a15ffa1af0c1224fcd888e47c`
`Hello, world! 12345` → `[19556, 28, 905, 17, 216, 33, 34, 35, 36, 37]`
` spaced text\n` → `[216, 23861, 216, 1694, 198]`
`\u4f60\u597d\uff0c\u4e16\u754c\uff01` → `[18645, 250, 48392, 138, 12831, 7906, 240, 178, 239, 230, 8083, 219]`
`Caf\xe9 \u2014 \u03ba\u03cc\u03c3\u03bc\u03bf\u03c2 \U0001f680` → `[51, 1939, 2756, 1841, 31953, 36180, 18751, 16674, 39346, 15107, 244, 218]`
`<\|endoftext\|>` → `[0]` |
| `talkie-13b-q4-native-tokenizer` | `PocketAiHub/talkie-1930-13b-it-GGUF@47b38329dd30e8b2d6ab8e2fc53f3f2ae789e694`
`talkie-1930-13b-it-Q4_K_M.gguf`
8,571,072,704 B
`2d6c6c1d98a1b8ffa38b50916454891a31ad844ee69c686e525976867917d7b2`
architecture `talkie`; declared pre `talkie`; effective pre `talkie` | `lewtun/talkie-1930-13b-it-hf@6311dedf518470856a8503f2080bb4b54fcb3323`
`config.json` 522 B `e7f29da9cf0a69571d6a0521cd912dc5c2f0dd151d0e934b87541f4389a9ee30`
`chat_template.jinja` 343 B `833a35215bfc10d1d9f27fb857123cc24bfef90f770fbc8d79ce37bf4ef4bc4d`, `tokenizer.json` 8,870,742 B `cc3813d9d674cf0e86e4171579ba276879c66c2171d993e5776fc5615756a03b`, `tokenizer_config.json` 247 B `e12d422a980eceaecd6ff388c3843b30dd461307d58ec19585953012d7386fc5` | validated identifiers `['talkie']`
metadata `7e14f443006afd16e49969f0bfbc5c995edde0075a829f2748e86b9fe4f2da81`
tokens 65,540 `f88816a5099baf479e674c8d3c61ed31f97954bd8213d1bf269cbbb883012b9e`
merges 65,279 `addf973bfd18babde5e7bfd7fe5f8e7fc3ae2f5fa01a7fdb4aba5c1898f0ec94`
types `511820f1cc6c9a5df9e0c93a062ff95f6d602208050aa12296fb66b798ad36cf`; scores=0
user-defined IDs `[]`; source added tokens=0 `None`
pipeline `{}`; chat `None`
source IDs `0..65539`; no GGUF-only padding extension; rows=65,540
materialized `63eb55af29f6eb88b2a8caa7966e0202b59f799d4e560bc688b5ac7c5f0453de`
`Hello, world! 12345` → `[72, 22882, 44, 1490, 33, 32, 6276, 1400]`
` spaced text\n` → `[32, 25156, 32, 5272, 10]`
`\u4f60\u597d\uff0c\u4e16\u754c\uff01` → `[228, 189, 160, 229, 165, 189, 239, 188, 140, 57632, 150, 231, 149, 140, 239, 188, 129]`
`Caf\xe9 \u2014 \u03ba\u03cc\u03c3\u03bc\u03bf\u03c2 \U0001f680` → `[67, 1063, 1238, 461, 12887, 12562, 6076, 7938, 14917, 32, 240, 159, 154, 128]`
`<\|user\|>hello<\|end\|>` → `[65537, 257, 12227, 65536]`
`<\|system\|>Be concise.<\|end\|><\|user\|>\u4f60\u597d 12345!<\|end\|><\|assistant\|>` → `[65539, 3664, 32185, 46, 65536, 65537, 228, 189, 160, 229, 165, 189, 32, 6276, 1400, 33, 65536, 65538]`
GGUF-native GPT4O reconstruction; official copy rejected: GGUF retains 65279 of 156379 source merges; first ordered mismatch at index 4 is ('Ġ', 'the') versus ('Ġt', 'he')
llama.cpp oracle `8d9af256337d1a501250f9bbf4c0859a654bddd6`: 444 cases `484246b629d6eec375ebac3672e4f4d4fb29646d3b331917ec4d2cfe385c3b6a` |
```python
from mobius.integrations.gguf import materialize_evidenced_gguf_tokenizer
materialize_evidenced_gguf_tokenizer("Qwen3.5-0.8B-Q4_0.gguf", "tokenizer")
```
Each row is independently artifact-scoped and proves ordered tokenizer semantics, source assets, embedding alignment, and the final materialized hash. Shared rows also require identical pinned llama.cpp dispatch. A matching complete immutable GGUF is automatically promoted to the pinned-source route during model and runtime package export; identifier-only inspection remains deferred because an identifier cannot prove artifact identity. This does not claim graph or runtime support.
### Fail-closed tokenizer evidence
- `glm4-7-flash-iq2-xxs-tokenizer-semantic-blocker` — **GGUF/source:** `bartowski/zai-org_GLM-4.7-Flash-GGUF@464d07505b441959737cd04d900f047469614c8d`
`zai-org_GLM-4.7-Flash-IQ2_XXS.gguf`
7,622,864,768 B
`b1f25d90e0da65587a5a8e359b40a9183c5a31b4908b3ee5ff370e05cc5e2ba4`
`zai-org/GLM-4.7-Flash@a9308079ef95921451a690cd2d16cb572e564642`
`chat_template.jinja` 3,120 B `d63ad536c3c81880043e22ec7fd08db42b4d8fb7c89c7138bc562bfa25281375`, `tokenizer.json` 20,217,442 B `19e773648cb4e65de8660ea6365e10acca112d42a854923df93db4a6f333a82d`, `tokenizer_config.json` 7,226 B `31a173e2797ddc8b72ac996803513e627fc28d7aad02cfcce321a431d865c86d`
`config.json` 1,070 B `dc9b97c7c9bed726a2e6939da4234d5c43abb3edec8812068c9a1af1dbc13acb`
first 9,475,456 B `803a3d88b31f81b5ac0fc541758af3c233d9d4a8e1abe7a601f2abaa52c5b382`; **closure:** architecture `deepseek2`; pre `glm4`
metadata `79115e0fa1bf5a8273b46eefb0e332e2b1e8957716311e8914fa76f7674d54ed`
tokens 154,880 `ba346b1893686f05f8e02b9a3657d4727b5bc8a2cecbd085bbe102c85fd683fe`; source 154,856 `6758484fa28bf2376d0eb319c51f0f5650b03156c3258b715ff031c9bc9f25f6`
merges 321,649 `7de84a5e88d436d1780d416b7c7c622800a556d691ff3c65b72c7aafcba3c47e`; source `7de84a5e88d436d1780d416b7c7c622800a556d691ff3c65b72c7aafcba3c47e`
scores=0; types `5bd4e8a8434ea2e20e413eb0484d1c3b93284bc05be3f3561331f621762b4e75`
added tokens `5284ad77c6affe20bd12e85e475c3906bcad5d0c02738de923cc0c4eba547cdb`; chat `d63ad536c3c81880043e22ec7fd08db42b4d8fb7c89c7138bc562bfa25281375`
normalizer `none`; pipeline `3c411b92af43f855284604c30bd7bb39f89ff83452f7f94c2a173f5c7c44c1f2`
exact aliases=['chatglm-bpe', 'glm4']; materialized `4ddc2c399cbef5431b5b1392a6e7417339b0ef7d096dcf2ba7658ef9327d333c`; config `d43221510295eff7c4afad3dfdbc8a0005e8a833fd543af4876458e2807fb132`
pipeline components={'decoder': '1d64d97add535d9ad91561aabea254849cf7f2ea4b924cc61c17152f1dd6e672', 'normalizer': '74234e98afe7498fb5daf1f36ac2d78acc339464f950703b8c019892f982b90b', 'post_processor': '18e6a4825c46b5a700a630cf00cb9329241348ac3b9e5eeb15b6b6936c165ffb', 'pre_tokenizer': '65e1fe6fbe22e0df7a3877257a5e19e7b0336324a8678a75de61b6fc0df0e147'}; added-token type mismatches=7; **witness:** fail-closed: ordered vocabulary, merges, deterministic padding, config, and chat template are pinned, but seven source added-token flags disagree with GGUF types; pinned CHATGLM4 splits apostrophe-space-CJK differently and detokenizes token 659 without the official leading space
`' \u597d`: llama.cpp `[6, 4891, 98, 121]` vs source `[6, 106992]`
corpus `4faf579ec80bff972046aa122be302fa1357e7ec6135fa94825b386493dfd09f`; llama.cpp oracle `8d9af256337d1a501250f9bbf4c0859a654bddd6`: 465 cases `2a821830ebc16fb14d8f1393002c0c17b84c3aa5af9e55765eeaec879dd6ab5f`
6 mismatches [2, 2, 2] by mode; source oracle `47d585793ea7cb663f36ff8ec48c2b474721e287044331bfe95a9eb50a0da3c2`
dispatch oracles={'chatglm-bpe': '2a821830ebc16fb14d8f1393002c0c17b84c3aa5af9e55765eeaec879dd6ab5f', 'glm4': '2a821830ebc16fb14d8f1393002c0c17b84c3aa5af9e55765eeaec879dd6ab5f'}; discriminator=('bailingmoe', 222, 'e7eaafadd6db83e5e8e33a418526181687797249baff65dace781dbb5bdbd7bb'); detokenize mismatches=33 [11, 11, 11] by mode; ` .`: llama.cpp hex `2e` vs source hex `202e`
- `llada-moe-iq1-s-tokenizer-semantic-blocker` — **GGUF/source:** `mradermacher/LLaDA-MoE-7B-A1B-Instruct-i1-GGUF@2ec29fbe69f07f382a864f93b40c4eecb45e6a0a`
`LLaDA-MoE-7B-A1B-Instruct.i1-IQ1_S.gguf`
1,717,318,112 B
`d711df4b4f819d9abd0e107469dd525eb12d3bc05ec173b6a8438c172f70f3de`
`inclusionAI/LLaDA-MoE-7B-A1B-Instruct@67004f662901b09f729994d4b3c04201283941ba`
`special_tokens_map.json` 153 B `f1fa4f8b8c24126a0c2a5d9b2de0fee32abbddf22f48c068e5cf42bc0a9b68ab`, `tokenizer.json` 7,663,358 B `4dd5931b0a63e3f61cfc1bcde132cd0c314de2f8a011ac9dbf2ff5efc40d0cbd`, `tokenizer_config.json` 4,593 B `ac03e164668db350d26b13bee7bc65fd7c4bc74595ac1fc3952ebd707b0b44a5`
`config.json` 1,424 B `59b6b803a1bf500b45249cb553b3fa0425e4f1a431ba9e7028ef6bda33c97586`
first 6,492,640 B `1d1fe0fcc1660d86157e99bc015910f14387a24029b9ba77820c674abdc9fc85`; **closure:** architecture `llada-moe`; pre `llada-moe`
metadata `e9b0281345bf2e3b859539f83f1738aab6552202285c437f36eee0539bc26d19`
tokens 157,184 `a31a2b2abd450d5750f4f28f507ff4d95efc1371b01a992b2374f9bb5fb59b53`; source 157,153 `8d89e7856d4a2b466cf0aee3ace415097cf77451c7e3a6a1f1b2fb2795485ad1`
merges 156,635 `0f44f5b5ed306c1f3617fbe217c8b3b3d382deb6129453fe3550e7dbf5453fac`; source `0f44f5b5ed306c1f3617fbe217c8b3b3d382deb6129453fe3550e7dbf5453fac`
scores=0; types `d6c379584304b9fe9ea5177810690cc6dd988960a224ca34f38d6ed08a6da10a`
added tokens `b65b415bd2707c17ee6e3e2e47b5be31c9e4972536482f4953f275e616f8beb4`; chat `a503d55b21709d881b07ee2e9d8ce29505fa95d9f758379a9fc4ccc94272734a`
normalizer `NFC`; pipeline `5401b80eb39f381efb73445a78a80b51652d34a05fca4d1fb3568b5fc473a0dd`
exact aliases=['bailingmoe', 'bailingmoe2', 'llada-moe']; materialized `702fd877edf627b1e01567071f161ceb96810cdd01afcf5e530b367547e757c9`; config `7b2b10ad497301177b1a0ab47bbcf938b17ca5e4b843cf45f62f01efda1851b7`
pipeline components={'decoder': '1d64d97add535d9ad91561aabea254849cf7f2ea4b924cc61c17152f1dd6e672', 'normalizer': '5628358406a1a2864e0f9e853618fc4d47d359acf10253558d0c18886de3f06e', 'post_processor': '18e6a4825c46b5a700a630cf00cb9329241348ac3b9e5eeb15b6b6936c165ffb', 'pre_tokenizer': '5e0cc01d5556750c556c081e65f2c0f9d9bbf8fa70dd98790cc9bfc5aafc7682'}; added-token type mismatches=0; **witness:** fail-closed: the exact GGUF and official source agree on ordered vocabulary prefix, merges, token types, added/special tokens, deterministic padding, config, and chat template, but the official NFC normalizer encodes decomposed e-acute as token 2900 while pinned BAILINGMOE preserves tokens 68 and 150766
`e\u0301`: llama.cpp `[68, 150766]` vs source `[2900]`
corpus `456ea43ace0bf9d548c6f5efdc7cc723832d3dd3fa7ad1d29f25d7dbf1aa210b`; llama.cpp oracle `8d9af256337d1a501250f9bbf4c0859a654bddd6`: 465 cases `a84708473c6afa75eb4b69fceb0f65c21e9fb57c8aea44a15cc227f8d565223d`
6 mismatches [2, 2, 2] by mode; source oracle `d30e15e4e8edfe0b51a9508081bb230fb58c1ad08b00b292396e7d4c719ace47`
dispatch oracles={'bailingmoe': 'a84708473c6afa75eb4b69fceb0f65c21e9fb57c8aea44a15cc227f8d565223d', 'bailingmoe2': 'a84708473c6afa75eb4b69fceb0f65c21e9fb57c8aea44a15cc227f8d565223d', 'llada-moe': 'a84708473c6afa75eb4b69fceb0f65c21e9fb57c8aea44a15cc227f8d565223d'}; discriminator=('glm4', 30, 'dec7509bc51a36ee3f60cbbaf03c08bb71b3cb15cbce167d87c9e163dbdbf6c1'); detokenize mismatches=6 [2, 2, 2] by mode; `e\u0301`: llama.cpp hex `65cc81` vs source hex `c3a9`
- `minicpm-2b-q2-k-tokenizer-mismatch` — **GGUF/source:** `mzwing/MiniCPM-2B-sft-bf16-GGUF@121e7290609857006939fca0ec64981009b806b9`
`MiniCPM-2B-sft-bf16.Q2_K.gguf`
1,204,392,288 B
`9e87235097895a22894c32a1e211f94b93b798d715cc7d99c8e846637927fd13`
`openbmb/MiniCPM-2B-sft-bf16@4ec16344ac13e6ef5010aeecaa533369ac8eb53c`
`special_tokens_map.json` 414 B `6fa06efa2785e450051989a6f8fb4416b10149ded485ddd3f127a40734f5cfd0`, `tokenizer.json` 6,202,715 B `42f73d01995bd71c88647b13ac696b36d84d5126d0c7cbaef6f8d872d5c97dff`, `tokenizer.model` 1,994,871 B `c9aafcd7da1f5611dab6be545db74d5552a2ccc9c2a12c72ea7be63aac4a25d7`, `tokenizer_config.json` 1,117 B `9c87efade54e9b26d3374a42f294d266722cbf9e97c748a12183d8317da074c9`
`config.json` 1,010 B `41cf26cfdca93f49209a6c0c26c00b281d6408e010ddc8cca9531c825cc17fc4`
first 16,777,216 B `ff140dcf42ce544e61a8e3bd23a09edbfb1a573d7df3f266081bcbd06e33ab66`; **closure:** architecture `minicpm`; pre `default`
metadata `1a596fac46038c46f2fa273fd8ec10cf0721acd2ed6daf2121cfc0840da9ee3e`
tokens 122,753 `154c42653082e70f4a98ed1708e8a71d41494dc1de32fe9316bcb6046b993dbe`; source 122,753 `154c42653082e70f4a98ed1708e8a71d41494dc1de32fe9316bcb6046b993dbe`
merges 0 `None`; source `4551998aaa2aa2468fc189317e85b7149a1999f38a9eddde19b523c3fa74e4b8`
scores=122753; types `e89c4d81e916e4d1bee6e7d7842463429004c559d9106d7e8af73934e9e4e0eb`
added tokens `8d90db0e62037d2ae62d998473452f7d104f8f5a03322d4589fc20d86da97602`; chat `ae9b050c5a5b0295cb09269e67bf832fa08675dd845c7b3ea4c2130bcacc5c26`
normalizer `Sequence(Prepend('\u2581'), Replace(' ', '\u2581'))`; pipeline `9cc85610fe636d4f21d99413627f6fcf6960886af506a3c537664ae01b421174`
source model tokens=122,753; source merges=171,540; score mismatches=122,752; type mismatches=1,087
GGUF chat `ae9b050c5a5b0295cb09269e67bf832fa08675dd845c7b3ea4c2130bcacc5c26` vs source `d9f25394f8be2d8a5fb234670a8596c7a12fa0d21896b5a757d77da8b8686944`; **witness:** fail-closed: all GGUF scores are -1000, GGUF merges are absent, token types and chat whitespace semantics diverge, and pinned llama.cpp disagrees with the official tokenizer on multilingual and whitespace inputs
`\u4f60\u597d\uff0c\u4e16\u754c\uff01`: llama.cpp `[29951, 95495, 65, 2925, 67]` vs source `[95320, 23523, 65, 2925, 67]`
corpus `dd6d2aa959ab8b74111c704559252bc8ef02aa5ec13f40ef23d0b65e03233413`; llama.cpp oracle `8d9af256337d1a501250f9bbf4c0859a654bddd6`: 444 cases `8959cbd62821def331adc77db30e8b351e8fad221547e74804dc7aecb188df1a`
18 mismatches [6, 6, 6] by mode; source oracle `a61293d6ad6fb5637f4ceafe225ac822897482d5ef5333b87f63b0e40611e9c2`
- `minicpm3-4b-q4-k-m-tokenizer-mismatch` — **GGUF/source:** `openbmb/MiniCPM3-4B-GGUF@816dc79b35f92827e0d2d87aacea3567e49661a8`
`minicpm3-4b-q4_k_m.gguf`
2,469,791,584 B
`64913247e927414ecf47fd3e9ea8e3f0c9acae293f583dfa7e24b8872e20fa4c`
`openbmb/MiniCPM3-4B@d6b14ddaefdb11c624dd75c3c779549bc90b08cb`
`added_tokens.json` 216 B `4760fcbf90bc193f33827ffe02f2e7ba1af1ec43644cc02ac22fdd611f6cca15`, `special_tokens_map.json` 1,632 B `068594063e37662c02b21acf42ebb334ef6a74fb810e68a2368f88f08351de76`, `tokenizer.json` 3,676,758 B `b00802b71a613e3f7df3899fe9643a3ff949736d333a2b892448a974383fe372`, `tokenizer.model` 1,181,204 B `bb74d51116831c3bf65db812c553f94ab0c88dcf97a5bbb37e3504f6d359c530`, `tokenizer_config.json` 10,413 B `25620d5a3f5727bba2fb403624f2c9a7bba55a7d00205829650cd1e3c646aae0`
`config.json` 1,929 B `cf1d08cb7c1815c676e685bd6ce94eb8b85a57d53871e6e159ee8c650717d98a`
first 16,777,216 B `505ce706e29108bef3579b1b4dc38695fe44923b63f453163175d5024b4ea12e`; **closure:** architecture `minicpm3`; pre `default`
metadata `6dc004393b6fd1dd27f81c505cadbd8be953244999f8b6fde281fb68dff94c34`
tokens 73,448 `1046ac4e64873087a848a2e033be381d18a35974f1ec0c139326c3073ad6744c`; source 73,448 `1046ac4e64873087a848a2e033be381d18a35974f1ec0c139326c3073ad6744c`
merges 0 `None`; source `a6ae9d2ba560703a2f5933b92307f80f9fdebfabc052df1cb5c3542a98441cbc`
scores=73448; types `ae14c00dfeff5f86796d58db7e91a7eee8fd44605d2cd3ec9846f56e52540a7e`
added tokens `b6fad720564107c65861ede9747fa08a03efba82c42268e963b4883abaf4f6d0`; chat `153280e3ff55d19da1398bdb3914ee2a51b80429bfaedde11d7d216c39db80f3`
normalizer `Sequence(Prepend('\u2581'), Replace(' ', '\u2581'))`; pipeline `9cc85610fe636d4f21d99413627f6fcf6960886af506a3c537664ae01b421174`
source model tokens=73,440; source merges=104,297; score mismatches=73,439; type mismatches=1,088
GGUF chat `153280e3ff55d19da1398bdb3914ee2a51b80429bfaedde11d7d216c39db80f3` vs source `dbd75fe18b14711fa5968600a6f5c974d7d3e63e75fe163ecb99a1e5f94c38c9`; **witness:** fail-closed: all GGUF scores are -1000, GGUF merges are absent, token types diverge, the GGUF drops the official tool-aware chat template, and pinned llama.cpp disagrees with the official tokenizer on multilingual and whitespace inputs
`\u4f60\u597d\uff0c\u4e16\u754c\uff01`: llama.cpp `[29951, 59495, 65, 2925, 67]` vs source `[59320, 23523, 65, 2925, 67]`
corpus `dd6d2aa959ab8b74111c704559252bc8ef02aa5ec13f40ef23d0b65e03233413`; llama.cpp oracle `8d9af256337d1a501250f9bbf4c0859a654bddd6`: 444 cases `d55374a7956f5379448b802ac23888c91f22aeb5fb6814f5b73efe74058fd475`
9 mismatches [3, 3, 3] by mode; source oracle `d6beee40e9257575b1333e43f67709920eda0cd6249fa74cfbd8a5eaa978b4af`
- `north-mini-code-iq1-s-tokenizer-semantic-blocker` — **GGUF/source:** `mradermacher/North-Mini-Code-1.0-i1-GGUF@94d8eb17eaeb728f907639ee0eff457e3e274667`
`North-Mini-Code-1.0.i1-IQ1_S.gguf`
6,455,984,128 B
`660792f0dd77ef2e39e92549bd88bbb0f91734371a3763816648fe77f23fb4dc`
`CohereLabs/North-Mini-Code-1.0@d11e61a842617a22dc328552fa5bb86231ee4f37`
`chat_template.jinja` 12,397 B `d8366efb9f07c571da620ce6a924594fc52c80273a0fbb46a38b643972df95fd`, `tokenizer.json` 28,217,141 B `14bd1c49d7d11874921d324986713df4be21cd06060530c497dacef99919b7a5`, `tokenizer_config.json` 8,954 B `1f45bd13ca86efccb5f74bf51a78c5e06f9066a5d4211499c7f81890f31d1da2`
`config.json` 2,342 B `0c987a88193e90c89a88a9dbeaba6844f5f24d00b728683338e2ace1476509a7`
first 10,428,416 B `9f38c617b8cd6fb3481cb1d73981cbbbd112e0e745b7cbd2e29c568ec11e76bb`; **closure:** architecture `cohere2moe`; pre `cohere2moe`
metadata `f164bb71b8a23b6fc0ca47163aa5effb62a6c7d3e92c1767590568ad0af7cbda`
tokens 262,144 `0d3ced2dd51c9cf8301847cf4204a7aa01d49fe23f83dca528f6736466d1b277`; source 255,032 `7ba11060fa3a8e3c527f75c69fed607660884a6d0ddafeba886dbae7695b538f`
merges 254,739 `79d571d309587b881fa56c5970d5de74ad6f3b4fde1af8a6356215ffc80d0fd6`; source `79d571d309587b881fa56c5970d5de74ad6f3b4fde1af8a6356215ffc80d0fd6`
scores=0; types `1c9ddca34d7f702a00b1f1795ebac3000980eb1104391e187243a49dc0cc4ead`
added tokens `62d98ff38dd8063a41e4d83a0282505044a654de200af9d020dce0422b386deb`; chat `d8366efb9f07c571da620ce6a924594fc52c80273a0fbb46a38b643972df95fd`
normalizer `none`; pipeline `68557134f65a45f8f2bb7e6dd6d74b6850a5ed3d1efdfa43c1d2f6c56500d329`
exact aliases=['cohere2moe', 'tiny_aya']; materialized `aa959ea5a543284ab7184a854c7f634223d3dfba40d2677f1078f8539303224b`; config `d33b45ace16bf59557c1e2b9b43d4bceb27d9f616304dfe59557abe7b0192e27`
pipeline components={'decoder': '1d64d97add535d9ad91561aabea254849cf7f2ea4b924cc61c17152f1dd6e672', 'normalizer': '74234e98afe7498fb5daf1f36ac2d78acc339464f950703b8c019892f982b90b', 'post_processor': '78dd5da36e83b02e49ef646d6bf6fcbba0f9dbd68cb9a4885929efd1725664cb', 'pre_tokenizer': '2778061f55c126b60e47da281317473fb373737f22ffe90419ed8efd7c9406cb'}; added-token type mismatches=23; **witness:** fail-closed: ordered vocabulary, merges, deterministic padding, config, and chat template are pinned, but 23 source added-token flags disagree with GGUF types and pinned TINY_AYA splits tab-space differently from the official North tokenizer
`\t 9`: llama.cpp `[202, 225, 29]` vs source `[13396, 29]`
corpus `ebd7b4f412ebf503c1ba171cdcd2aef3b9e5a4a28e3fd7a8e6f391be57df54f5`; llama.cpp oracle `8d9af256337d1a501250f9bbf4c0859a654bddd6`: 465 cases `c4736b65ca7e9b609b269068990842fa3538f2fc76081db010e464fe0cdb7c86`
14 mismatches [6, 4, 4] by mode; source oracle `c56b02c040f04640c56f0821b2beb694d4570005e42ba4ceb47502ef38ac717f`
dispatch oracles={'cohere2moe': 'c4736b65ca7e9b609b269068990842fa3538f2fc76081db010e464fe0cdb7c86', 'tiny_aya': 'c4736b65ca7e9b609b269068990842fa3538f2fc76081db010e464fe0cdb7c86'}; discriminator=('glm4', 69, '814ab562958b77eaa6dbffc4e8ae2883c5ffd259f96c411ca2912c60574ab5cc'); detokenize mismatches=0 [0, 0, 0] by mode
- `plm-1.8b-instruct-q4-k-m-tokenizer-blocker` — **GGUF/source:** `PLM-Team/PLM-1.8B-Instruct-gguf@7bec6546983bcf0d99526c943580bd49e2237445`
`PLM-1.8B-Instruct-Q4_K_M.gguf`
1,182,708,992 B
`b38570ee56ebec82a1e9ef45ab408c0d8230ececef1d7f1b267c49cff35638b8`
`PLM-Team/PLM-1.8B-Instruct@62d188c7d58843d7013d5b3ffe198db448787860`
`merges.txt` 1,671,853 B `8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5`, `special_tokens_map.json` 410 B `c83747485fba9ef20c42793b4b02b05001f214250f0d787f573df216c91047a3`, `tokenizer.json` 11,418,266 B `bcfe42da0a4497e8b2b172c1f9f4ec423a46dc12907f4349c55025f670422ba9`, `tokenizer_config.json` 1,327 B `1becffcfa09c98935043f1724d988887c618c5f6e7a249087d3ae29eb70e2a6f`, `vocab.json` 2,776,833 B `ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910`
`config.json` 934 B `91e6e13695a6de82556438667e64b60d9910269f300cd97f8c667d19e75f115e`; **closure:** architecture `plm`; pre `qwen2`
metadata `698bfa31cd069292437bc3509fea7be2445324a536d95a6e813d947c283bd989`
tokens 151,936 `f3ea8e8cf45bd58a8d5ad420306a3ccd925894cdb61a89062fd9e3a6de255a0e`; source 151,646 `696f26322524de87f49427fd1be6d1afce910574d9656c5d3f4f64064bdb83c1`
merges 151,387 `24fa2ae2a398e50784a1fff678482094af4f63e6783d35686726abacda8dc371`; source `cc098baa4a74ce5156487605aa048a34e54f0eee6a704691a738c8fb22dafdd5`
scores=0; types `0286431feb975d95a59a3f39957f8183d929295635fb627b6940627c63918bf1`
added tokens `e7b5f7013431aa26739424d92f40423f175a46f6d1fdc8453edf6005c99412f7`; chat `af9c0233881b083b52ff773580215222b5440ac3d0beeeca99b76329b048f8db`
normalizer `NFC`; pipeline `97c53ee89fb584b10798f44b02c60c9a8b746165a32dc34737d178fc20618a69`; **witness:** official tokenizer.json applies NFC normalization, but pinned llama.cpp qwen2 preserves decomposed Unicode; exact materialization is blocked
`e\u0301 \xe9`: llama.cpp `[68, 53839, 3958]` vs source `[963, 3958]`
corpus `0f3c77d24208f2ac0833668128cb2a00adcb7e6b4b4eedc6e4783e3ec7b41f5a`; llama.cpp oracle `8d9af256337d1a501250f9bbf4c0859a654bddd6`: 72 cases `66513168812575ccac974ecb454e916def5f4492d558c1866b8811d4f587a41d`
The MiniCPM, Gemma4, and final alias-group fixtures are reproducible through `scripts/generate_*tokenizer*.py`, which validates immutable bounded headers and official tokenizer hashes, builds tokenizer-only GGUFs and the pinned llama.cpp helper, then recomputes exact outputs and mismatch witnesses. Committed Gemma4 and alias-group inputs replay materialized identities network-free; the alias oracle never calls the production reconstruction.
## Supported GGUF architectures
Reason codes are concise user-facing categories; detailed architecture audits remain in
`_arch_registry.py` and its tests.
| Canonical architecture | Aliases | Import route | Tensor exactness | Config/tensor/graph/runtime/quantized import | Restriction or evidence gap |
|---|---|---|---|---|---|
| `afmoe` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — AFMoE combines sandwich norms, Q/K norms, sigmoid-gated attention, MuP embedding scaling, a dense prefix, correction-biased routed/shared experts, and optional interleaved sliding-window attention. |
| `apertus` | — | model=`apertus`; tensor=`llama`+`apertus_extras` | exact-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=supported; quantized_import=supported | EVIDENCED_SCOPE — Runtime support is restricted to the pinned Apertus-v1.1-1.5B-Instruct BF16 artifact's exact-float CPU route, official tokenizer revision, full-logit stateful evidence, and ORT GenAI 0.15.2. |
| `arcee` | — | model=`arcee`; tensor=`arcee` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `arctic` | — | model=`arctic`; module=`arctic_gguf`; tensor=`llama`+`arctic_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — The mobius graph uses floating Linear modules for this architecture, so no MatMulNBits or BlockQuantizedMatMul target can consume preserved GGUF projection weights. |
| `arwkv7` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — ARWKV7 wraps RWKV7's delta-rule matrix recurrence in a distinct one-shift RMSNorm/Qwen residual topology with optional five-versus-six-way interpolation, optional gate/group norm, and Qwen SwiGLU. |
| `baichuan` | — | model=`baichuan`; tensor=`llama` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `bailingmoe` | — | model=`bailing_moe`; tensor=`llama`+`diffusion_fused_qkv`+`moe_extras` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `bailingmoe2` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — BailingMoE2 serializes complete dense-or-routed/shared expert trailing blocks plus NextN and layer-output norms, but the pinned loader marks every trailing tensor skipped and exposes no MTP graph. |
| `bailingmoe3` | — | none (fails before config extraction) | not claimed | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — BailingMoE3 alternates head-wise KDA recurrent layers with gated MLA layers, so each sequence carries three causal-convolution histories plus a matrix state alongside attention cache. |
| `bert` | — | model=`bert`; tensor=`bert` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact pinned tensor closure, encoder-only task dispatch, and synthetic ORT execution are covered, but no pinned real GGUF artifact has passed independent embedding parity. |
| `bitnet` | — | model=`bitnet`; tensor=`bitnet` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — The exact float route is implemented: TQ1_0/TQ2_0 blocks are dequantized by the pinned gguf reader, optional scalar projection scales are folded into those values, and the dedicated graph owns both sub-norms. |
| `bloom` | — | model=`bloom`; tensor=`bloom` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Quantization preservation is rejected because canonical Bloom GGUF stores one fused QKV projection that must be reordered and split into three graph targets. |
| `chameleon` | — | none (fails before config extraction) | exact-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — The pinned Chameleon converter deliberately omits the VQ image tokenizer while the text graph still requires bias-bearing Q/K norms, an additional swin_norm, and image-vocabulary logit suppression. |
| `chatglm` | — | model=`chatglm`; tensor=`chatglm`; mmproj=`generic_projector` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Quantization preservation is rejected because fused QKV and gate/up tensors must be split into separate packed graph targets. |
| `clip` | — | none (fails before config extraction) | not claimed | config=rejected; tensor_map=rejected; graph=rejected; runtime=rejected; quantized_import=rejected | CONFIG_REJECTED — This is a multimodal projector sidecar, not a language model. |
| `codeshell` | — | model=`kclgpt`; tensor=`legacy_layernorm` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Quantization preservation is rejected because the pinned loader accepts a fused QKV tensor that must be split into separate graph projections. |
| `cogvlm` | — | none (fails before config extraction) | exact-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — The pinned CogVLM text graph has modality-routed visual-expert Q/K/V/output and FFN banks in addition to the language projections. |
| `cohere2` | — | model=`cohere2`; tensor=`llama` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `cohere2moe` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — Cohere2MoE's executable head uses sigmoid routed fused-or-split experts, optional shared experts, no FFN norm, and interleaved sliding-window KV state. |
| `command-r` | — | model=`command_r`; tensor=`llama`+`command_r_extras` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Import requires canonical logit_scale metadata and is restricted to split Q/K/V tensors in the 40-layer Command-R profile; quantization preservation is supported only for that split route. |
| `dbrx` | — | model=`dbrx`; module=`dbrx_gguf`; tensor=`llama`+`dbrx_extras`+`moe_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — The mobius graph uses floating Linear modules for this architecture, so no MatMulNBits or BlockQuantizedMatMul target can consume preserved GGUF projection weights. |
| `deci` | — | model=`llama`; tensor=`llama` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `deepseek` | — | model=`deepseek`; tensor=`llama`+`diffusion_fused_qkv`+`deepseek_shared_moe_extras` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `deepseek2` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — DeepSeek2 MTP is a complete MLA plus routed/shared MoE block with compressed KV cache, Q/KV LoRA alternatives, target-owned embedding/head fallbacks, and architecture-specific gating. |
| `deepseek2-ocr` | — | none (fails before config extraction) | exact-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — DeepSeek-OCR2 is a paired text-plus-vision package, not a generic DeepSeek text model. |
| `deepseek32` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — DeepSeek3.2 extends the DeepSeek2 MLA/MoE head with DSA indexer projections, normalization, bias, and sparse-cache metadata. |
| `deepseek4` | — | none (fails before config extraction) | not claimed | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — The pinned DeepSeek-V4 GGUF runtime uses a dedicated raw sliding-window, CSA, HCA, and indexer compressed-cache ABI with persistent compressor state, rollback snapshots, four-stream hyper-connections, hash/sqrt-softplus routing, and optional MTP storage. |
| `dflash` | — | model=`DFlashDraftModel`; tensor=`dflash` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | DIRECT_ORT_EVIDENCED / RUNTIME_UNVALIDATED — Exact target-coupled direct ORT acceptance, rollback, and deterministic generation are evidenced; higher-level runtime compatibility remains advisory. |
| `dots1` | — | model=`dots1`; tensor=`llama`+`diffusion_fused_qkv`+`moe_qk_norm_extras`+`deepseek_shared_moe_extras` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `dots3note` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — Dots3Note preserves an MLA/DSA trunk and a dense sliding-MLA NextN block, but the pinned loader explicitly has no MTP graph and skips the head. |
| `dream` | — | model=`dream`; tensor=`llama`+`diffusion_fused_qkv` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, suffix-exact tensor closure, masked-diffusion task dispatch, and synthetic full-sequence execution are covered, but no pinned real GGUF has passed independent Hugging Face/llama.cpp masked-step logit parity and deterministic multi-step generation parity. |
| `eagle3` | — | model=`Eagle3DraftModel`; tensor=`eagle3` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | DIRECT_ORT_EVIDENCED / RUNTIME_UNVALIDATED — Exact target-coupled direct ORT acceptance, rollback, and deterministic generation are evidenced; higher-level runtime compatibility remains advisory. |
| `ernie4_5` | — | model=`ernie4_5`; module=`gguf_legacy`; tensor=`legacy_layernorm` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Import is narrowed to the dense split-Q/K/V, split-SwiGLU, full-RoPE variant and rejects all expert, fused, sectioned-position, and bias alternatives. |
| `ernie4_5-moe` | — | model=`ernie4_5_moe`; module=`ernie4_5_moe_gguf`; tensor=`llama`+`moe_extras`+`ernie45_moe_extras` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — The mobius graph uses floating Linear modules for this architecture, so no MatMulNBits or BlockQuantizedMatMul target can consume preserved GGUF projection weights. |
| `eurobert` | — | model=`eurobert`; module=`eurobert_gguf`; tensor=`eurobert` | exact-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Config extraction, exact pinned tensor closure, encoder-only task dispatch, and synthetic ORT execution are covered, but no pinned real GGUF artifact has passed independent embedding parity. |
| `exaone` | — | model=`exaone`; tensor=`llama` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `exaone-moe` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — EXAONE-MoE serializes a dense trailing NextN block after an iSWA routed/shared expert trunk, but the pinned loader skips appended blocks. |
| `exaone4` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — EXAONE4 serializes attention/FFN post-norm trailing blocks and NextN tensors with optional synthetic Llama3 RoPE factors, but the pinned loader skips them. |
| `falcon` | — | model=`falcon`; tensor=`falcon` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `falcon-h1` | — | model=`falcon_h1`; tensor=`falcon_h1` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — The dedicated graph and GGUF importer preserve parallel Attention+Mamba2 layers and their four-state ABI. |
| `gemma` | — | model=`gemma`; tensor=`llama` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `gemma-embedding` | — | model=`gemma3_text`; module=`gemma_embedding_gguf`; tensor=`gemma_embedding` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — The exact float embedding graph and package ABI are covered, but no onnxruntime-genai embedding package schema/runtime evidence is recorded. |
| `gemma2` | — | model=`gemma2`; tensor=`llama`+`gemma2_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `gemma3` | — | model=`gemma3_text`; tensor=`llama`+`gemma3_extras`; mmproj=`gemma3` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `gemma3n` | — | none (fails before config extraction) | exact-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — Gemma3n GGUF is the text member of a vision-and-audio package whose gemma3nv and gemma3na clip companions carry distinct encoders and projectors. |
| `gemma4` | — | model=`gemma4_text`; tensor=`llama`+`gemma4_extras`; mmproj=`gemma4` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `gemma4-assistant` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — Gemma4 Assistant is a standalone target-coupled model with pre/post projections, masked embeddings, scalar layer scales, its own KV cache, and a live target-model context. |
| `glm-dsa` | `glm_dsa` | model=`glm_moe_dsa`; tensor=`glm_dsa` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Exact explicit-float MLA, DSA indexer, dense-prefix/routed-MoE tensor routing, and packed dynamic-cache graph construction are covered. |
| `glm4` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — GLM4 serializes complete fused-FFN trailing blocks and NextN tensors, but the pinned loader skips appended blocks; GLM-OCR converter transforms also permute Q/K for M-RoPE. |
| `glm4moe` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — GLM4-MoE serializes biased attention and periodic dense/routed expert trailing blocks with mandatory router bias, but the pinned loader skips them. |
| `gpt-oss` | — | none (fails before config extraction) | not claimed | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — The pinned GPT-OSS converter splits interleaved gate/up expert rows and repacks checkpoint block+scale tensors into expert-major MXFP4 values. |
| `gpt2` | — | model=`gpt2`; tensor=`gpt2` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=supported; quantized_import=rejected | FLOAT_IMPORT_ONLY — Runtime support is restricted to the exact evidenced GPT-2 Q2_K artifact, explicit-float CPU import route, pinned tokenizer, and ORT GenAI 0.15.2. |
| `gptj` | — | none (fails before config extraction) | no-loader | config=rejected; tensor_map=rejected; graph=rejected; runtime=rejected; quantized_import=supported | CONFIG_REJECTED — The pinned census reserves gptj but llama.cpp has no model loader for it. |
| `gptneox` | — | model=`gpt_neox`; module=`gguf_legacy`; tensor=`legacy_layernorm` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=supported; quantized_import=rejected | FLOAT_IMPORT_ONLY — Runtime support is restricted to the exact evidenced Pythia-70M Q2_K artifact, explicit-float portable graph, pinned tokenizer, and ORT GenAI 0.15.2. |
| `granite` | — | model=`granite`; tensor=`llama`+`diffusion_fused_qkv`+`moe_extras` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Exact float and quantization-preserving import covers the pinned dense-or-MoE union, fused or split QKV, optional projection biases, ungated shared experts, and Granite scaling. |
| `granite_swa` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — Granite SWA requires attention sinks, a complete interleaved sliding-window schedule, residual/logit scaling, fused routed gate-up experts, and optional fused shared experts/deep-stack injection. |
| `granitehybrid` | — | model=`granitemoehybrid`; tensor=`granitehybrid` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Exact mixed attention/Mamba2 scheduling, architecture-wide dense or routed MoE feed-forward selection, optional shared experts, Granite scaling, value-preserving float expert fusion, and strict pinned tensor closure are supported. |
| `granitemoe` | — | model=`granitemoe`; tensor=`llama`+`moe_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `graniteswitch` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — GraniteSwitch repurposes an appended synthetic layer as a token-history-driven adapter router and carries fourteen switched-LoRA tensors per block in addition to decoder KV state. |
| `grok` | — | model=`grok_gguf`; module=`grok_gguf`; tensor=`grok` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — The mobius graph uses floating Linear modules for this architecture, so no MatMulNBits or BlockQuantizedMatMul target can consume preserved GGUF projection weights. |
| `grovemoe` | — | model=`grovemoe_gguf`; module=`grovemoe_gguf`; tensor=`grovemoe` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — The mobius graph uses floating Linear modules for this architecture, so no MatMulNBits or BlockQuantizedMatMul target can consume preserved GGUF projection weights. |
| `hunyuan-dense` | `hunyuan_v1_dense` | model=`hunyuan_v1_dense`; tensor=`llama`+`hunyuan_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `hunyuan-moe` | — | model=`hunyuan_moe_gguf`; module=`hunyuan_moe_gguf`; tensor=`hunyuan_moe` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — The mobius graph uses floating Linear modules for this architecture, so no MatMulNBits or BlockQuantizedMatMul target can consume preserved GGUF projection weights. |
| `hunyuan_vl` | — | none (fails before config extraction) | exact-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — The pinned Hunyuan-VL decoder uses its own M-RoPE and Q/K-normalized text contract and pairs with a hunyuanvl clip sidecar. |
| `hy_v3` | — | model=`hy_v3`; tensor=`llama`+`hy_v3_extras` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Exact full-attention trunk and independently cached NextN sidecar graphs are covered, including per-head Q/K RMSNorm, a contiguous dense prefix, selection-biased sigmoid routed SwiGLU experts, and an ungated shared expert. |
| `internlm2` | — | model=`internlm2`; tensor=`llama` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — The mobius graph uses floating Linear modules for this architecture, so no MatMulNBits or BlockQuantizedMatMul target can consume preserved GGUF projection weights. |
| `jais` | — | model=`jais`; module=`gguf_legacy`; tensor=`legacy_layernorm` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Converter-baked MuP scales are retained exactly, while fused biased QKV must be split and therefore cannot preserve packed quantization. |
| `jais2` | — | model=`jais2`; tensor=`legacy_layernorm` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `jamba` | — | model=`jamba`; tensor=`jamba` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Exact mixed attention/Mamba and dense/routed-MoE schedules, strict tensor closure and shapes, GGUF value transforms, compatible projection quantization, value-checked expert ordering, reduced Transformers parity, and multi-token ORT state threading, reorder, and replay are covered. |
| `jina-bert-v2` | — | model=`bert`; module=`jina_bert_v2_gguf`; tensor=`jina_bert_v2` | exact-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Config extraction, exact pinned tensor closure, encoder-only task dispatch, and synthetic ORT execution are covered, but no pinned real GGUF artifact has passed independent embedding parity. |
| `jina-bert-v3` | — | model=`jina-bert-v3`; module=`jina_bert_v3_gguf`; tensor=`jina_bert_v3` | exact-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — The dedicated graph and float importer preserve the pinned RoPE, post-norm, and sequential GELU contracts of the reachable dense loader path. |
| `kimi-k3` | — | model=`kimi_k3`; tensor=`kimi_k3` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — The exact KDA/NoPE gated-MLA schedule, four-state recurrent ABI, attention-residual banks, Stable LatentMoE routing, SiTU activation, strict metadata/tensor closure, and compatible projection quantization are supported. |
| `kimi-linear` | — | model=`kimi_linear`; tensor=`kimi_linear` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — The exact KDA/NoPE-MLA schedule, four-state recurrent ABI, dense/MoE topology, correction-bias routing, pinned metadata, tensor closure, and compatible MatMul/expert quantization are supported. |
| `laguna` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — Laguna combines per-head-or-element softplus attention gates, dual-RoPE interleaved sliding-window attention, a dense prefix, and sigmoid correction-biased routed/shared experts. |
| `lfm2` | — | model=`lfm2`; tensor=`lfm2` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=supported; quantized_import=supported | EVIDENCED_SCOPE — Runtime support is restricted to LiquidAI's official LFM2-350M F16 artifact, pinned CPU import route, exact tokenizer revision, hybrid convolution/KV state evidence, and ORT GenAI 0.15.2. |
| `lfm2moe` | — | model=`lfm2_moe`; tensor=`lfm2`+`lfm2_moe_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Config extraction, exact pinned tensor-name closure, GGUF value transforms, and synthetic recurrent-state execution are covered, but no representative real-weight GGUF has yet passed independent full-logit parity and deterministic multi-token stateful ORT generation. |
| `llada` | — | model=`llada`; tensor=`llama` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, suffix-exact tensor closure, masked-diffusion task dispatch, and synthetic full-sequence execution are covered, but no pinned real GGUF has passed independent Hugging Face/llama.cpp masked-step logit parity and deterministic multi-step generation parity. |
| `llada-moe` | — | model=`llada`; module=`llada_moe`; tensor=`llama`+`diffusion_fused_qkv`+`moe_qk_norm_extras`+`moe_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, suffix-exact tensor closure, masked-diffusion task dispatch, and synthetic full-sequence execution are covered, but no pinned real GGUF has passed independent Hugging Face/llama.cpp masked-step logit parity and deterministic multi-step generation parity. |
| `llama` | `mistral` | model=`llama`; tensor=`llama`; mmproj=`generic_projector` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=supported; quantized_import=supported | EVIDENCED_SCOPE — Runtime support is restricted to exact structured evidence matches. |
| `llama-embed` | — | model=`llama`; module=`llama_embed_gguf`; tensor=`llama_embedding` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Only the complete dense, split-QKV, unbiased, default-RoPE float profile is promoted. |
| `llama4` | — | none (fails before config extraction) | exact-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — Llama4 GGUF is the text member of a paired multimodal package and may contain routed experts and architecture-specific cross-modal layer scheduling. |
| `maincoder` | — | model=`maincoder`; tensor=`llama`+`exact_legacy_gguf_extras` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Exact float import is covered with learned per-head Q/K RMSNorm after adjacent-pair RoPE, sequential pre-norm SwiGLU blocks, causal GQA cache, and a tied output head. |
| `mamba` | — | model=`mamba`; tensor=`mamba` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Config extraction, exact pinned tensor-name closure, GGUF value transforms, and synthetic recurrent-state execution are covered, but no representative real-weight GGUF has yet passed independent full-logit parity and deterministic multi-token stateful ORT generation. |
| `mamba2` | — | model=`mamba2`; tensor=`mamba2` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Config extraction, exact pinned tensor-name closure, GGUF value transforms, and synthetic recurrent-state execution are covered, but no representative real-weight GGUF has yet passed independent full-logit parity and deterministic multi-token stateful ORT generation. |
| `mellum` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — Mellum requires an untied head, Q/K norms, routed experts in every layer, and a metadata-defined full/sliding attention schedule with distinct RoPE behavior. |
| `mimo2` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — MiMo2 requires fused-QKV dense MTP blocks, attention sinks, interleaved sliding KV cache, and three chained heads selected by offsets. |
| `minicpm` | — | model=`minicpm`; module=`minicpm_gguf`; tensor=`llama`; mmproj=`generic_projector` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Exact dense graph import, tensor closure, scaling, LongRoPE factors, Q/K transforms, and expanded dynamic K/V cache are covered. |
| `minicpm3` | — | model=`minicpm3`; module=`minicpm3_gguf`; tensor=`minicpm3` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Exact Q/KV-LoRA MLA graph import with expanded K/V cache is covered. |
| `minimax-01` | — | model=`minimax`; tensor=`minimax` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Graph import is exact, but released ORT GenAI packaging cannot represent the heterogeneous KV/recurrent state slots or bounded rollback snapshots; runtime packaging remains tracked by #605. |
| `minimax-m2` | — | model=`minimax_m2_gguf`; module=`minimax_m2_gguf`; tensor=`minimax_m2` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Exact explicit-float import owns MiniMax-M2's non-square Q/O geometry, full-vector Q/K RMSNorm, partial NeoX RoPE, standard dynamic/static KV cache, and F32 selection-biased sigmoid MoE routing. |
| `minimax-m3` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — MiniMax-M3 adds F32 sparse-indexer tensors and a second index-key cache with position/cell maps, block masks, rollback, and reorder semantics alongside main K/V state. |
| `mistral3` | — | none (fails before config extraction) | exact-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — The pinned Mistral3 loader selects dense or routed-expert text blocks from metadata and applies architecture-specific output temperature scaling. |
| `mistral4` | — | model=`mistral4_gguf`; module=`mistral4_gguf`; tensor=`mistral4` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Exact explicit-float import owns Mistral4's DeepSeek-V2 MLA projections, dense-prefix plus mandatory shared/routed MoE blocks, YaRN scaling, and one graph-visible latent-plus-RoPE K cache per layer. |
| `modern-bert` | — | model=`modernbert`; tensor=`modern_bert` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact pinned tensor closure, encoder-only task dispatch, and synthetic ORT execution are covered, but no pinned real GGUF artifact has passed independent embedding parity. |
| `mpt` | — | model=`mpt`; module=`gguf_legacy`; tensor=`legacy_layernorm` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=supported; quantized_import=rejected | FLOAT_IMPORT_ONLY — Runtime support is restricted to the exact evidenced tiny MPT Q2_K artifact, explicit-float portable graph, pinned tokenizer, and ORT GenAI 0.15.2. |
| `muse-glimmer` | `muse_glimmer` | model=`muse_glimmer_text`; tensor=`llama`+`muse_glimmer_extras`; mmproj=`muse_glimmer` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `nanbeige` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — Nanbeige reuses physical layer weights across a configurable logical loop count, optionally normalizes between loops, and allocates a distinct KV slot for every logical occurrence. |
| `nemotron` | — | model=`nemotron`; tensor=`llama` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `nemotron_h` | — | model=`nemotron_h`; tensor=`nemotron_h` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Config extraction, exact pinned tensor-name closure, GGUF value transforms, and synthetic recurrent-state execution are covered, but no representative real-weight GGUF has yet passed independent full-logit parity and deterministic multi-token stateful ORT generation. |
| `nemotron_h_moe` | — | model=`nemotron_h`; tensor=`nemotron_h_moe` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Exact mixed attention/Mamba2/dense/MoE scheduling, sigmoid correction-bias routing, shared experts, optional latent projections, and strict GGUF tensor closure are supported. |
| `neo-bert` | — | model=`neobert`; module=`neo_bert_gguf`; tensor=`neo_bert` | exact-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Config extraction, exact pinned tensor closure, encoder-only task dispatch, and synthetic ORT execution are covered, but no pinned real GGUF artifact has passed independent embedding parity. |
| `nomic-bert` | — | model=`nomic_bert`; module=`nomic_bert_gguf`; tensor=`nomic_bert` | exact-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Config extraction, exact pinned tensor closure, encoder-only task dispatch, and synthetic ORT execution are covered, but no pinned real GGUF artifact has passed independent embedding parity. |
| `nomic-bert-moe` | — | model=`nomic_bert_moe`; module=`nomic_bert_moe_gguf`; tensor=`nomic_bert`+`nomic_bert_moe_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Config extraction, exact pinned tensor closure, encoder-only task dispatch, and synthetic ORT execution are covered, but no pinned real GGUF artifact has passed independent embedding parity. |
| `olmo` | — | model=`olmo`; tensor=`olmo` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=supported; quantized_import=supported | EVIDENCED_SCOPE — Runtime support is restricted to the exact evidenced tiny OLMo Q2_K artifact, explicit-float CPU import route, pinned tokenizer, and ORT GenAI 0.15.2. |
| `olmo2` | — | model=`olmo2`; tensor=`llama`+`olmo2_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `olmoe` | — | model=`olmoe`; tensor=`llama`+`moe_qk_norm_extras`+`moe_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `openelm` | — | model=`openelm`; module=`gguf_legacy`; tensor=`legacy_layernorm`+`exact_legacy_gguf_extras` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Quantization preservation is rejected because every OpenELM layer stores fused QKV rows that must be split into per-layer Q/K/V graph projections. |
| `orion` | — | model=`orion`; tensor=`legacy_layernorm` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Fused QKV input is rejected because its import transform is not implemented. |
| `paddleocr` | — | none (fails before config extraction) | strongest-converter-family-inventory-loader-inherited-from-ernie4_5-with-optional-attn-output-bias | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — PaddleOCR-VL uses an ERNIE-derived GGUF loader with an optional bias on attention output closure and a required paddleocr clip vision/projector sidecar. |
| `pangu-embedded` | — | model=`pangu_embedded`; tensor=`llama` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Import is intentionally restricted to split Q/K/V tensors and ordinary full-head RoPE. |
| `phi2` | — | model=`phi`; tensor=`phi2` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Quantization preservation is rejected because the Phi-2 attention, MLP, and output graph uses float-only linear modules. |
| `phi3` | — | model=`phi3`; tensor=`phi3` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `phimoe` | — | model=`phimoe`; module=`phimoe_gguf`; tensor=`llama`+`phi3`+`moe_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `plamo` | — | model=`plamo`; module=`gguf_plamo`; tensor=`plamo` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Exact graph/config/tensor import is implemented for the pinned PLaMo-13B converter contract. |
| `plamo2` | — | model=`plamo2`; tensor=`plamo2` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — The dedicated graph and GGUF importer preserve PLaMo2's alternating Mamba1/attention layers and mixed state ABI. |
| `plamo3` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — PLaMo3 requires fused QKV and fused SwiGLU, four norm sites with architecture-specific offset transforms, Q/K norm before RoPE, and a periodic full/sliding attention state (seven sliding layers then one full layer by default). |
| `plm` | — | model=`plm`; tensor=`plm` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `pockettts` | — | none (fails before config extraction) | not claimed | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — The primary GGUF is only PocketTTS's transformed causal CALM backbone: its embedding table contains folded learned conditioning rows and its duplicated embedding output is not a semantic LM head. |
| `qwen` | — | model=`qwen`; tensor=`llama`+`qwen1_extras` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Quantization preservation is rejected because Qwen v1 stores fused QKV weights that must be split into separate graph projections. |
| `qwen2` | — | model=`qwen2`; tensor=`llama` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=supported; quantized_import=supported | EVIDENCED_SCOPE — Runtime support is restricted to the official Qwen2.5-0.5B-Instruct Q8_0 artifact, pinned CPU import route, exact tokenizer revision, and ORT GenAI 0.15.2 evidence. |
| `qwen2moe` | `qwen2_moe` | model=`qwen2_moe`; tensor=`llama`+`moe_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `qwen2vl` | — | model=`qwen2_vl_text`; tensor=`llama`; mmproj=`qwen_vl` | exact-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Text and paired Qwen2/Qwen2.5-VL projector graph import are supported for the exact split-QKV llama.cpp artifacts, but downstream multimodal runtime execution has not been evidenced. |
| `qwen3` | — | model=`qwen3`; tensor=`llama`+`moe_qk_norm_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `qwen35` | — | model=`qwen3_5_text`; tensor=`llama`+`qwen35_hybrid_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact pinned tensor-name closure, GGUF value transforms, and synthetic recurrent-state execution are covered, but no representative real-weight GGUF has yet passed independent full-logit parity and deterministic multi-token stateful ORT generation. |
| `qwen35moe` | — | model=`qwen3_5_moe`; tensor=`llama`+`moe_extras`+`qwen35_hybrid_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=supported; quantized_import=supported | EVIDENCED_SCOPE — Runtime support is restricted to the pinned Q2_K reduced checkpoint's explicit-float CPU route, exact tokenizer revision, hybrid KV/convolution/recurrent-state evidence, and ORT GenAI 0.15.2. |
| `qwen3moe` | `qwen3_moe` | model=`qwen3_moe`; tensor=`llama`+`moe_qk_norm_extras`+`moe_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `qwen3next` | — | model=`qwen3_next`; tensor=`llama`+`moe_extras`+`qwen3next_hybrid_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact pinned tensor-name closure, GGUF value transforms, and synthetic recurrent-state execution are covered, but no representative real-weight GGUF has yet passed independent full-logit parity and deterministic multi-token stateful ORT generation. |
| `qwen3tts` | — | none (fails before config extraction) | not claimed | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — The primary GGUF is only a transformed Qwen3-TTS talker backbone, not the existing Mobius Qwen3TTS conditional-generation or codec model. |
| `qwen3vl` | — | none (fails before config extraction) | exact-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — Qwen3-VL text GGUF requires multimodal position IDs and an exact qwen3vl_merger clip companion, including deep-stack vision features and architecture-specific token placement. |
| `qwen3vlmoe` | — | none (fails before config extraction) | exact-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — Qwen3-VL-MoE combines the Qwen3-VL multimodal position/token contract and merger sidecar with routed experts in the text backbone. |
| `qwen4exp` | — | header/config/tensor preflight only; model=`qwen4_exp_text`; tensor=`qwen4exp` | exact pinned 3-shard artifact header closure | config=supported; tensor_map=supported; graph=deferred; runtime=deferred; quantized_import=rejected | GRAPH_DEFERRED — Exact header/config/tensor preflight is implemented, but no executable graph import route is claimed. |
| `refact` | — | model=`refact`; module=`gguf_legacy`; tensor=`legacy_layernorm` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Import is narrowed to split, bias-free dense tensors with one KV head; loaded-but-unexecuted expert, RoPE-factor, and bias families are rejected. |
| `rnd1` | — | model=`llada`; module=`rnd1`; tensor=`llama`+`diffusion_fused_qkv`+`moe_qk_norm_extras`+`moe_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, suffix-exact tensor closure, masked-diffusion task dispatch, and synthetic full-sequence execution are covered, but no pinned real GGUF has passed independent Hugging Face/llama.cpp masked-step logit parity and deterministic multi-step generation parity. |
| `rwkv6` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — RWKV6 carries two F32 states per layer (two token-shift vectors and a per-head WKV matrix) and applies token-dependent exp(-exp(decay)), a time_first read-before-update term, per-head group norm, and cumulative rescale transforms. |
| `rwkv6qwen2` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — RWKV6-Qwen2 is neither Qwen2 attention nor native RWKV6: it carries one F32 token-shift vector plus a per-head matrix state and uses k*(1-w) gated linear attention, optional biased/GQA projections, a sigmoid gate, and parallel Qwen SwiGLU. |
| `rwkv7` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — RWKV7 requires a two-shift F32 state plus a per-head matrix state, generalized delta-rule recurrence, six-way token mixing, first-layer value residuals shared across depth, ICLR/key-adaptation vectors, and an r_k residual around LayerNorm and group norm. |
| `seed_oss` | — | model=`seed_oss`; tensor=`llama`+`seed_oss_extras` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `smallthinker` | — | model=`smallthinker_gguf`; module=`smallthinker_gguf`; tensor=`smallthinker` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — The exact float-import graph owns SmallThinker's pre-norm router, ReGLU experts, metadata-selected sigmoid/softmax gate, and per-layer SWA/NoPE schedule. |
| `smollm3` | — | model=`smollm3`; tensor=`llama` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `stablelm` | — | model=`stablelm`; tensor=`llama` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
| `starcoder` | — | model=`gpt_bigcode`; tensor=`starcoder` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=supported; quantized_import=rejected | FLOAT_IMPORT_ONLY — Runtime support is restricted to the exact evidenced tiny StarCoder Q2_K artifact, explicit-float CPU import route, pinned tokenizer, and ORT GenAI 0.15.2. |
| `starcoder2` | — | model=`starcoder2`; tensor=`llama` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=supported; quantized_import=supported | EVIDENCED_SCOPE — Runtime support is restricted to the exact evidenced tiny StarCoder2 Q2_K artifact, explicit-float CPU import route, pinned tokenizer, and ORT GenAI 0.15.2. |
| `step35` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — Step3.5 executes one or more interleaved-SWA NextN heads with optional gates, routed/shared experts, centered-norm transforms, per-layer head geometry, and dedicated cache offsets. |
| `t5` | — | model=`t5`; tensor=`t5` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Graph import is covered, but no independent full-logit and generation parity run has yet validated a pinned real T5 GGUF runtime package. |
| `t5encoder` | — | model=`t5encoder`; tensor=`t5` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Encoder hidden-state import is covered, but the pinned real artifact lacks independent provenance and full hidden-state parity evidence. |
| `talkie` | — | model=`talkie`; tensor=`talkie` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — The exact pinned scalar-sidecar float graph is covered, including weight-free RMSNorm, inverse NeoX RoPE, post-RoPE Q/K normalization, per-head Q gain, embedding skip, causal KV cache, and logit scaling. |
| `wavtokenizer-dec` | — | none (fails before config extraction) | not claimed | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — wavtokenizer-dec is a stateless non-causal code-token to ISTFT-parameter network, not a waveform codec decoder. |
| `xverse` | — | model=`xverse`; tensor=`llama` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=rejected | RUNTIME_EVIDENCE_PENDING / FLOAT_IMPORT_ONLY — Fused QKV input is rejected because it cannot be combined truthfully with the required architecture-specific Q/K row permutations. |
## Stored quantization types
The generated machine-readable source for this table is
`testdata/evidence/gguf_quantization_capabilities.json`. It records parse and exact
dequantization support separately from conversion, names the implementation transform and
operator ABI for every tensor role, and treats dequantize/requantize as non-preserving.
| Stored qtype | ID | Parse | Exact dequantization | Projection/output route | Direct exactness | Embedding route | Expert-major route | Target storage | Source fidelity | Native operator ABI | Runtime evidence |
|---|---:|---|---|---|---|---|---|---|---|---|---|
| `Q4_0` | 2 | supported | supported | affine repack | exact | affine repack | affine repack | quantized target supported | true | — | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `Q4_1` | 3 | supported | supported | affine repack | lossy | affine repack | affine repack | quantized target supported | false | — | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `Q5_0` | 6 | supported | supported | dequantize/requantize | — | dequantize/requantize | dequantize/requantize | quantized target supported | false | — | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `Q5_1` | 7 | supported | supported | dequantize/requantize | — | dequantize/requantize | dequantize/requantize | quantized target supported | false | — | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `Q8_0` | 8 | supported | supported | affine repack | exact | affine repack | affine repack | quantized target supported | true | — | `qwen2.5-0.5b-instruct-q8-ort-genai-0.15.2` |
| `Q2_K` | 10 | supported | supported | dequantize/requantize | — | dequantize/requantize | dequantize/requantize | quantized target supported | false | — | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `Q3_K` | 11 | supported | supported | dequantize/requantize | — | dequantize/requantize | dequantize/requantize | quantized target supported | false | — | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `Q4_K` | 12 | supported | supported | affine repack | lossy | affine repack | affine repack | quantized target supported | false | — | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `Q5_K` | 13 | supported | supported | dequantize/requantize | — | dequantize/requantize | dequantize/requantize | quantized target supported | false | — | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `Q6_K` | 14 | supported | supported | affine repack | lossy | affine repack | affine repack | quantized target supported | false | — | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `IQ2_XXS` | 16 | supported | supported | native byte-preserved | — | dequantize/requantize | native byte-preserved | quantized target supported | true | `pkg.nxrt::BlockQuantizedMatMul/v1` (`iq2_xxs`) | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `IQ2_XS` | 17 | supported | supported | native byte-preserved | — | dequantize/requantize | native byte-preserved | quantized target supported | true | `pkg.nxrt::BlockQuantizedMatMul/v1` (`iq2_xs`) | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `IQ3_XXS` | 18 | supported | supported | native byte-preserved | — | dequantize/requantize | native byte-preserved | quantized target supported | true | `pkg.nxrt::BlockQuantizedMatMul/v1` (`iq3_xxs`) | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `IQ1_S` | 19 | supported | supported | native byte-preserved | — | dequantize/requantize | native byte-preserved | quantized target supported | true | `pkg.nxrt::BlockQuantizedMatMul/v1` (`iq1_s`) | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `IQ4_NL` | 20 | supported | supported | native byte-preserved | — | dequantize/requantize | native byte-preserved | quantized target supported | true | `pkg.nxrt::BlockQuantizedMatMul/v1` (`iq4_nl`) | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `IQ3_S` | 21 | supported | supported | native byte-preserved | — | dequantize/requantize | native byte-preserved | quantized target supported | true | `pkg.nxrt::BlockQuantizedMatMul/v1` (`iq3_s`) | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `IQ2_S` | 22 | supported | supported | native byte-preserved | — | dequantize/requantize | native byte-preserved | quantized target supported | true | `pkg.nxrt::BlockQuantizedMatMul/v1` (`iq2_s`) | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `IQ4_XS` | 23 | supported | supported | native byte-preserved | — | dequantize/requantize | native byte-preserved | quantized target supported | true | `pkg.nxrt::BlockQuantizedMatMul/v1` (`iq4_xs`) | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `IQ1_M` | 29 | supported | supported | native byte-preserved | — | dequantize/requantize | native byte-preserved | quantized target supported | true | `pkg.nxrt::BlockQuantizedMatMul/v1` (`iq1_m`) | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `TQ1_0` | 34 | supported | supported | dequantize/requantize | — | dequantize/requantize | dequantize/requantize | quantized target supported | false | — | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `TQ2_0` | 35 | supported | supported | dequantize/requantize | — | dequantize/requantize | dequantize/requantize | quantized target supported | false | — | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `MXFP4` | 39 | supported | supported | native byte-preserved | — | dequantize/requantize | native byte-preserved | quantized target supported | true | `pkg.nxrt::BlockQuantizedMatMul/v1` (`mxfp4`) | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `NVFP4` | 40 | supported | supported | dequantize/requantize | — | dequantize/requantize | dequantize/requantize | quantized target supported | false | — | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `Q1_0` | 41 | supported | deferred | affine repack | exact | affine repack | affine repack | quantized target supported | true | — | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
| `Q2_0` | 42 | supported | deferred | rejected | — | rejected | rejected | rejected | false | — | deferred: No real-weight ONNX Runtime execution evidence is recorded. |
## Multimodal projector sidecars
| Artifact ID | Immutable sidecar | Bytes | SHA-256 | Projector types | Paired text | Processor source |
|---|---|---:|---|---|---|---|
| `gemma3n-e4b-f16` | `Qwe1325/gemma-3n-E4B-it-GGUF@f26cfdb3f7e86ede704fc45410316e48ccb1a018`
`mmproj-F16.gguf` | 1,967,809,568 | `a464216b97121e8065216569fc501880cc456ef859e41234235106b7348e0279` | `gemma3nv`, `gemma3na` | `Qwe1325/gemma-3n-E4B-it-GGUF@f26cfdb3f7e86ede704fc45410316e48ccb1a018`
`gemma-3n-E4B-it-q3_k_m.gguf`
3,442,332,096 bytes | `unsloth/gemma-3n-E4B-it@45e9fb1dd0e34db5ff9db1f43a49ac5d8e8b8778` |
| `gemma4-unified-12b-f16` | `unsloth/gemma-4-12b-it-GGUF@fc034cfff751157913579611efad8462ac1be606`
`mmproj-F16.gguf` | 175,115,840 | `91f086971e56d7a7d8d39e271873fccdb49541bd259d6e02c401a4f1cb7a219e` | `gemma4uv`, `gemma4ua` | `unsloth/gemma-4-12b-it-GGUF@fc034cfff751157913579611efad8462ac1be606`
`gemma-4-12b-it-Q4_K_M.gguf`
7,121,861,440 bytes | `google/gemma-4-12B-it@707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7` |
| `smolvlm-256m-idefics3-f16` | `ggml-org/SmolVLM-256M-Instruct-GGUF@b9e4379657e1450d04d02eec8e345667265b0a00`
`mmproj-SmolVLM-256M-Instruct-f16.gguf` | 190,031,616 | `0802360aca1748f112ea510b8ff277c65b1361c8ef30ed89b83c9c7a60d08e96` | `idefics3` | `ggml-org/SmolVLM-256M-Instruct-GGUF@b9e4379657e1450d04d02eec8e345667265b0a00`
`SmolVLM-256M-Instruct-f16.gguf`
327,809,728 bytes | `HuggingFaceTB/SmolVLM-256M-Instruct@7e3e67edbbed1bf9888184d9df282b700a323964` |
| `internvl25-1b-f16` | `ggml-org/InternVL2_5-1B-GGUF@d77253530c9a27486a28800afaf6ff5576c0bf17`
`mmproj-InternVL2_5-1B-f16.gguf` | 619,876,960 | `0c672edd99ec0b99df01c75dbb6cc26ad2236d7d61f908c93b5fda9b4d9ddd20` | `internvl` | `ggml-org/InternVL2_5-1B-GGUF@d77253530c9a27486a28800afaf6ff5576c0bf17`
`InternVL2_5-1B-f16.gguf`
1,265,481,408 bytes | `OpenGVLab/InternVL2_5-1B@9d423ea1ae9f893897ee3f7493141073f5afcf22` |
| `llama4-scout-f16` | `ggml-org/Llama-4-Scout-17B-16E-Instruct-GGUF@42675345da11ade9203a5187595da7b74d4ff2ac`
`mmproj-Llama-4-Scout-17B-16E-Instruct-f16.gguf` | 1,746,780,608 | `a7eec12068ae70f993fbba6eb350c095727be20f7a6ecbe6e431940c1a8823fb` | `llama4` | `complete Llama4 Scout text GGUF exceeds 16 GiB` | `meta-llama/Llama-4-Scout-17B-16E-Instruct@92f3b1597a195b523d8d9e5700e57e4fbb8f20d3` |
| `pixtral-12b-f16` | `ggml-org/pixtral-12b-GGUF@cba1ea4420bc2b4f15f50fdec59e30769880a63c`
`mmproj-pixtral-12b-f16.gguf` | 870,070,176 | `b4819558d6524a2e5623a06104ee085253a6dfd2b51470c60771ec33976f81bb` | `pixtral` | `ggml-org/pixtral-12b-GGUF@cba1ea4420bc2b4f15f50fdec59e30769880a63c`
`pixtral-12b-Q2_K.gguf`
4,791,047,808 bytes | `mistral-experimental/pixtral-12b@c2756cbbb9422eba9f6c5c439a214b0392dfc998` |
| `llava-llama3-8b-mlp-f16` | `xtuner/llava-llama-3-8b-v1_1-gguf@344f1bfe987bcbdc7e650b134d23670d5ffb5892`
`llava-llama-3-8b-v1_1-mmproj-f16.gguf` | 624,434,368 | `eb569aba7d65cf3da1d0369610eb6869f4a53ee369992a804d5810a80e9fa035` | `mlp` | `bartowski/Meta-Llama-3-8B-Instruct-GGUF@4ebc4aa83d60a5d6f9e1e1e9272a4d6306d770c1`
`Meta-Llama-3-8B-Instruct-Q2_K.gguf`
3,179,131,456 bytes | `xtuner/llava-llama-3-8b-v1_1-transformers@b20fb3040caaf5d0b3751c0d86a94efdf5bb007d` |
| `mobilevlm-1.7b-ldp-f16` | `guinmoon/MobileVLM-1.7B-GGUF@7e0cdbd2d642d938ce82fadde991360500c7d7cf`
`MobileVLM-1.7B-mmproj-f16.gguf` | 620,384,896 | `7d9855d323cee2a1797a88f9d7057ce26b21dcd62a50b382c4ff44ea60c77e39` | `ldp` | `guinmoon/MobileVLM-1.7B-GGUF@7e0cdbd2d642d938ce82fadde991360500c7d7cf`
`MobileVLM-1.7B-Q4_K.gguf`
834,055,776 bytes | — |
| `mobilevlm-v2-1.7b-ldpv2-f16` | `ZiangWu/MobileVLM_V2-1.7B-GGUF@422c888cc387d71831bedf48d59f0a66b27fad68`
`mmproj-model-f16.gguf` | 595,103,072 | `57966afa654e9d46a11b2a4b17989c2d487cd961f702c4fe310f86db5e30aab4` | `ldpv2` | `ZiangWu/MobileVLM_V2-1.7B-GGUF@422c888cc387d71831bedf48d59f0a66b27fad68`
`ggml-model-q4_k.gguf`
791,817,856 bytes | — |
| `glm-edge-v-2b-adapter-f16` | `zai-org/glm-edge-v-2b-gguf@d76cbe14f1d3a9405f664cbb5ae0c9537197429a`
`mmproj-model-f16.gguf` | 933,229,600 | `69a11ec5f54219fef9fd6bf9bc3209f0e6ef1564462cc4705dd93b2cd2a8198c` | `adapter` | `zai-org/glm-edge-v-2b-gguf@d76cbe14f1d3a9405f664cbb5ae0c9537197429a`
`ggml-model-Q4_0.gguf`
931,269,056 bytes | `THUDM/glm-edge-v-2b@2053707733f99ab52e943904f43c2359a94301ef` |
| `minicpm-v2-resampler-f16` | `openbmb/MiniCPM-V-2-gguf@3a38804c39d96c935a6b542581f51171aefa06a5`
`mmproj-model-f16.gguf` | 866,071,872 | `79611c59b5ad5b0547256602e3fb546a3041bcf6db5058091b6bcaa31f3a1c95` | `resampler` | `openbmb/MiniCPM-V-2-gguf@3a38804c39d96c935a6b542581f51171aefa06a5`
`ggml-model-Q2_K.gguf`
1,297,193,376 bytes | — |
| `qwen2-vl-2b-f16` | `ggml-org/Qwen2-VL-2B-Instruct-GGUF@bb307c036e8a1ed7b663bbd0c35b41c4c9294cfd`
`mmproj-Qwen2-VL-2B-Instruct-f16.gguf` | 1,331,656,160 | `ecb20cabcdd8dbc277de06bd6eb980aeb2adfaaba9f199a434e328d205675d03` | `qwen2vl_merger` | `Qwen2-VL-2B-Instruct-Q4_K_M.gguf` | `Qwen/Qwen2-VL-2B-Instruct@895c3a49bc3fa70a340399125c650a463535e71c` |
| `qwen25-vl-3b-f16` | `ggml-org/Qwen2.5-VL-3B-Instruct-GGUF@5037fcf163dd95d1e41d1974465f0898ed108ca2`
`mmproj-Qwen2.5-VL-3B-Instruct-f16.gguf` | 1,338,428,128 | `b9160fe9d814d1fadf68395677468534778b39ac33c2e7561b7b218626e60d5e` | `qwen2.5vl_merger` | `Qwen2.5-VL-3B-Instruct-Q4_K_M.gguf` | `Qwen/Qwen2.5-VL-3B-Instruct@66285546d2b821cf421d4f5eb2576359d3770cd3` |
| `gemma3-4b-f16` | `ggml-org/gemma-3-4b-it-GGUF@ab31416aceb30cd095cb34cc27eea120940964e4`
`mmproj-model-f16.gguf` | 851,251,104 | `8c0fb064b019a6972856aaae2c7e4792858af3ca4561be2dbf649123ba6c40cb` | `gemma3` | `gemma-3-4b-it-Q4_K_M.gguf` | `google/gemma-3-4b-it@093f9f388b31de276ce2de164bdc2081324b9767` |
| `gemma4-e2b-f16` | `unsloth/gemma-4-E2B-it-GGUF@0314792d7f1f7e229411f620751375812bb9faf2`
`mmproj-F16.gguf` | 985,654,080 | `337ee849e80b6169ce9d1d573d424fc1653bcafa5f0cb0cbb901beba54f4b41c` | `gemma4v`, `gemma4a` | `unsloth/gemma-4-E2B-it-GGUF@0314792d7f1f7e229411f620751375812bb9faf2`
`gemma-4-E2B-it-Q4_K_M.gguf`
3,106,738,272 bytes | `google/gemma-4-E2B-it@3e22461f65e89153144f8adb70e3b8c2cc9845a7` |
| `muse-glimmer-30b-bf16` | `unsloth/Muse-Glimmer-30B-GGUF@faa5b025c584459c13febfa5c59883516710ae39`
`mmproj-Muse-Glimmer-30B-BF16.gguf` | 3,849,173,728 | `7aa788cfe25ae5e4bf4837511f64df22cabe595e58223708274a67b3136f53ab` | `muse-glimmer` | `Muse-Glimmer-30B-UD-Q4_K_XL.gguf` | — |
| `qwen3-vl-projector-f16` | `bartowski/Qwen_Qwen3-VL-2B-Instruct-GGUF@e84f8ae7ffee8b04793a4ed771609e2b61d3f3cf`
`mmproj-Qwen_Qwen3-VL-2B-Instruct-f16.gguf` | 819,394,848 | `8c3f6a56979a1ce7056b9a20be6cf6b6f6ad4837aa3da532b5afcfcfd1faa38b` | `qwen3vl_merger` | `Qwen_Qwen3-VL-2B-Instruct-Q4_K_M.gguf` | `Qwen/Qwen3-VL-2B-Instruct@89644892e4d85e24eaac8bacfd4f463576704203` |
| `qwen3-audio-projector-bf16` | `ggml-org/Qwen3-ASR-0.6B-GGUF@928ab958557df9aa2ef1c93e0e83c7ad0933fae2`
`mmproj-Qwen3-ASR-0.6B-bf16.gguf` | 378,575,520 | `dae36c855f9a82a8916bea2238b24bda69a39d8da8b2f46dee7c103775656039` | `qwen3a` | `Qwen3-ASR-0.6B-bf16.gguf` | `Qwen/Qwen3-ASR-0.6B@5eb144179a02acc5e5ba31e748d22b0cf3e303b0` |
| `qwen2-audio-projector-f16` | `mradermacher/Qwen2-Audio-7B-Instruct-GGUF@e1e68850ba33e38eafbc3817919c318d9c7e757b`
`Qwen2-Audio-7B-Instruct.mmproj-f16.gguf` | 1,289,301,536 | `b52435dead2956f1fc113818c3b5ceb42a940cb487e59163cb1ffc69cae69347` | `qwen2a` | `Qwen2-Audio-7B-Instruct.Q4_K_M.gguf` | `Qwen/Qwen2-Audio-7B-Instruct@0a095220c30b7b31434169c3086508ef3ea5bf0a` |
| `qwen25-omni-projector-f16` | `ggml-org/Qwen2.5-Omni-3B-GGUF@75f1b73b657a50f5092502799457ccb4a4a1f9df`
`mmproj-Qwen2.5-Omni-3B-f16.gguf` | 2,623,983,328 | `f6d9276e9fa4f060c7abdbe886786cf31a8911b62770f8a54b7581b7b99fa27e` | `qwen2.5o`, `qwen2.5vl_merger`, `qwen2a` | `Qwen2.5-Omni-3B-Q4_K_M.gguf` | `Qwen/Qwen2.5-Omni-3B@f75b40e3da2003cdd6e1829b1f420ca70797c34e` |
| `glm4v-projector-f16` | `mradermacher/GLM-OCR-GGUF@3c1e642c0fa5df64831f0b04f3c674b57ce341af`
`GLM-OCR.mmproj-f16.gguf` | 869,018,080 | `fe5805b3b70f3174d25a912b8d197569eaa8e1e3e6d9777a385b8cc4c622af6c` | `glm4v` | `GLM-OCR.Q4_K_M.gguf` | `zai-org/GLM-OCR@ca5d8b3e287e52589e37c28385d9655ee4372f9d` |
| `ultravox-v0.5-f16` | `ggml-org/ultravox-v0_5-llama-3_2-1b-GGUF@5390c7c41cbd6f261f7f205fc0c5ae61bbdca650`
`mmproj-ultravox-v0_5-llama-3_2-1b-f16.gguf` | 1,371,123,616 | `b34dde1835752949d6b960528269af93c92fec91c61ea0534fcc73f96c1ed8b2` | `ultravox` | `fixie-ai/ultravox-v0_5-llama-3_2-1b` | `fixie-ai/ultravox-v0_5-llama-3_2-1b@b95bec8ab291eeb04b5cd600dd473377f6b79026` |
| `music-flamingo-bf16` | `henry1477/music-flamingo-gguf@a059053433697011c6928b1962110040f4bcb4d0`
`mmproj-music-flamingo-bf16.gguf` | 1,324,506,912 | `d4be69ed65f25dae97062febd44f9f41c0f6b14178f1cfb530fd894d595a4f94` | `musicflamingo` | `nvidia/music-flamingo-hf` | `nvidia/music-flamingo-hf@35a2c9071753ee075b0f7fc2fd81151c21389530` |
| `lfm2.5-audio-1.5b-f16` | `LiquidAI/LFM2.5-Audio-1.5B-GGUF@7d525f883a077e20afb782f2ff618edcae0e39e4`
`mmproj-LFM2.5-Audio-1.5B-F16.gguf` | 458,806,624 | `71330d7820768417d950f2dce42227896c7f6146917453957a63ba765decf621` | `lfm2a` | `LiquidAI/LFM2.5-Audio-1.5B` | `LiquidAI/LFM2.5-Audio-1.5B@c362a0625dfe45aa588dce5f0ada28a7e5707628` |
| `granite-speech-4.1-2b-f16` | `ibm-granite/granite-speech-4.1-2b-GGUF@8267dad2adc84209b0efd2702ec68a98356125eb`
`mmproj-model-f16.gguf` | 1,159,354,752 | `0d3615076cbe1d35c3f60c43a60a4047b3e2eeee1b2c233580be60186faab5c5` | `granite_speech` | `ibm-granite/granite-speech-4.1-2b` | `ibm-granite/granite-speech-4.1-2b@de575db64086f84fdc79da4932d1076e965bc546` |
| `cogvlm-chat-v1.1-f16-header` | `PandaExpressPatron/cogvlm-chat-gguf@8076a26b5563f170569805cf17e22401f5c790e8`
`mmproj-cogvlm-chat-hf`
first 16,777,216 B `f06cb0cfd4974db17f9549e87178dbfba65146ecf2529b35c2badb09ed4909b6` | 8,858,987,936 | `aa2e53f40d8248738fc79f790227fde7f564bce870e496db89b5e90425fa1a4b` | `cogvlm` | `cogvlm-13B-chat-v1.1-F16.gguf` | `zai-org/cogvlm-chat-hf@e29dc3ba206d524bf8efbfc60d80fc4556ab0e3c` |
| `exaone4-5-33b-f16-header` | `LGAI-EXAONE/EXAONE-4.5-33B-GGUF@0e969634ef24db05151b435970297a6dee634b7e`
`mmproj-EXAONE-4.5-33B-F16.gguf`
first 16,777,216 B `2aa263be72bbb57706da85fd14515b6e1c323028ee620d68aa8bb7f6d5d1a5f3` | 2,574,221,920 | `98d2ecab5b64edea00314150f276ac4de5816b980dbf591832efc344e0f6295a` | `exaone4_5` | `EXAONE-4.5-33B-IQ4_XS.gguf` | `LGAI-EXAONE/EXAONE-4.5-33B@570aa4b15a4f45ba1133072b45f50198f6e3b4fd` |
| `hunyuanocr-bf16-header` | `ggml-org/HunyuanOCR-GGUF@8e070c9ad79e4ca97a9b4daa2f1ce17e8759afb1`
`mmproj-HunyuanOCR-bf16.gguf`
first 16,777,216 B `15f375226c21c2f069f00abd9eddca211347c93f323bc09df50b859e6b88f086` | 997,235,840 | `46401739a91d0778d86369bb952db685b215512d61a941c3b859f337f6014fcd` | `hunyuanvl` | `HunyuanOCR-bf16.gguf` | `tencent/HunyuanOCR@b7bf72439f11fa076c547edf8777aa85f8e0a027` |
| `janus-pro-1b-f16-header` | `mradermacher/Janus-Pro-1B-GGUF@31ced8c1d0bd842eeb8f27730ba792b16e7cda91`
`Janus-Pro-1B.mmproj-f16.gguf`
first 16,777,216 B `039c75d5059a38bfa2274da7c8b1511c4a9d642c84e23cede104beac33f4d226` | 621,824,384 | `b1f441fa6ef80e9b058808e100afeb65f2a844736b8f4865791b59fad1453b1c` | `janus_pro` | `Janus-Pro-1B.Q8_0.gguf` | `deepseek-community/Janus-Pro-1B@1655280bb75959cc1cb85529a2a8b26e7016072e` |
| `kimi-k2-5-f16-header` | `AesSedai/Kimi-K2.5-GGUF@43ea7b530645c4f1d2616ec0d376d92b6e69cc9f`
`mmproj-Kimi-K2.5-F16.gguf`
first 16,777,216 B `3e7c895e916044cadd5faa3b68ab6b023ec2d33ca98013d8d98efa71ec3a5ad2` | 952,572,160 | `9261f190d7b8561fc69f70d2bbbc533d5975704f19c6d2b08fa8ac6c133ec78c` | `kimik25` | `Kimi-K2.5 text GGUF split set` | `moonshotai/Kimi-K2.5@4d01dfe0332d63057c186e0b262165819efb6611` |
| `kimi-vl-a3b-f16-header` | `ggml-org/Kimi-VL-A3B-Thinking-2506-GGUF@e7dcd093335f922a057772febc7ab27eda985b40`
`mmproj-Kimi-VL-A3B-Thinking-2506-f16.gguf`
first 16,777,216 B `fe078c1ecf7a1b00ebf3e040d54754b3fc32ef0500c75c53c5df29f199ff7289` | 905,371,712 | `1386854a5031970a92ec458ddcc2dfce575f28edf6463643e50ac2f008068fb3` | `kimivl` | `Kimi-VL-A3B-Thinking-2506-Q4_K_M.gguf` | `moonshotai/Kimi-VL-A3B-Thinking-2506@aa1730989e7558695b44ee493623e03bd325a994` |
| `lfm2-vl-1-6b-f16-header` | `LiquidAI/LFM2-VL-1.6B-GGUF@6121de267003bb4d4f325fe10abdc735aee06747`
`mmproj-LFM2-VL-1.6B-F16.gguf`
first 16,777,216 B `6744dec15d156b007be7a12adbd667a4442ee956f0fcbddbe53e5c195cf00149` | 830,339,008 | `b637bfa6060be2bc7503ec23ba48b407843d08c2ca83f52be206ea8563ccbae2` | `lfm2` | `LFM2-VL-1.6B-Q4_0.gguf` | `LiquidAI/LFM2.5-VL-1.6B@919fde3d022e3f90a4716006f993938ee8c2eb97` |
| `mimo-v2-5-f16-header` | `AesSedai/MiMo-V2.5-GGUF@eed9c5e5d55c5cf9fba5309e66ad5246ea3ffa13`
`mmproj-MiMo-V2.5-F16.gguf`
first 16,777,216 B `4e0adf899ed6ad0e485107ce6ecfd38e89da6e688b3006a2787ea3e912f64392` | 1,458,190,112 | `e1664b236f6fef3d3c64abb24d4441580cbe0df34333995afc1de8ef96ef5089` | `mimovl` | `MiMo-V2.5 text GGUF split set` | `XiaomiMiMo/MiMo-V2.5@63651580ca774f8504f676040460aed3e1244ac1` |
| `minicpm-v4-6-bf16-header` | `prithivMLmods/MiniCPM-V-4.6-GGUF@01a7c0ceb731b733bfe9dc3875dc08004f3596b4`
`MiniCPM-V-4.6.mmproj-bf16.gguf`
first 16,777,216 B `964d497bc9097b54837f410bd2b04c069256d1835e2a9a7efcfda082fa2aac60` | 1,110,101,888 | `7b6ffe05cfbc8afbeec7f1847c865c9b6ca1d78f5e128ff7136536b4a87b67e7` | `minicpmv4_6` | `MiniCPM-V-4.6.Q4_K_M.gguf` | `openbmb/MiniCPM-V-4.6@36f34a661a4bd35d0dc2294cb044d2584646c7d3` |
| `nemotron-nano-v2-vl-bf16-header` | `tomlawrence/NVIDIA-Nemotron-Nano-12B-v2-VL-GGUF@2b87641797d12e8a316ab509c7752887a6d2660a`
`mmproj-BF16.gguf`
first 16,777,216 B `33c5e9f28a5d7499dd7fc72058df5ee9915a08dda68cf5dc991a3b1d37e8d9d4` | 1,689,151,936 | `10f5fdb09ae2e122af388b5905c4921e66d8f4cc8382f6d1152fcfe8b3fa229d` | `nemotron_v2_vl` | `NVIDIA-Nemotron-Nano-12B-v2-VL-Q4_0.gguf` | `nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16@ca9543b126e8bf3176916d3d305ccc415f89fd4d` |
| `step3-vl-10b-f16-header` | `JamePeng2023/Step3-VL-10B-GGUF@f0bc308167bb03e463e10ba8dfaa4879092d0e84`
`mmproj-Step3-VL-10b-F16.gguf`
first 16,777,216 B `eb458306d565cd859d4e3db2cc802c1b92138d80b5b4bc9bb6233cf692648753` | 3,972,829,344 | `541f08bdb3f4799a31a7d37708ba7990fa173bdb14b233c1485076afe34a08a9` | `step3vl` | `Step3-VL-10B-Q3_K_M.gguf` | `stepfun-ai/Step3-VL-10B@5026053b0c2f5dfaa08fc2d149384162c3c8bca1` |
| `yasa2-reka-edge-f16-header` | `Vastined/reka-edge-2603-GGUF@e6a3fd4d8012aea11c7b0d54b3af020bd2e34366`
`mmproj-reka-edge-2603-F16.gguf`
first 16,777,216 B `fef11ed9c78e6a7a3b79bcfe58678ddef6bfc7c29d3924351d9fdc5e96034ec5` | 1,376,166,016 | `2671b816c9dc63c075fb41615afb7881a7f5e3e7969c874767573ce5a6cd7bc8` | `yasa2` | `reka-edge-2603-Q4_K_M.gguf` | `RekaAI/reka-edge-2603@492c81c225fbf5f3263a8245b00827721b119a13` |
| `deepseek-ocr-bf16` | `sabafallah/DeepSeek-OCR-GGUF@d26779bcd1cb301fec3ff82adc672f18384776fc`
`mmproj-deepseek-ocr-bf16.gguf` | 826,425,472 | `4caeed8b6c3c7d25dfebccfdb5cf34d6ae540ef4dc4fa2b9842b69cfa50ecbe2` | `deepseekocr` | `sabafallah/DeepSeek-OCR-GGUF@d26779bcd1cb301fec3ff82adc672f18384776fc`
`deepseek-ocr-Q4_K_M.gguf`
1,950,326,592 bytes | `deepseek-ai/DeepSeek-OCR@9f30c71f441d010e5429c532364a86705536c53a` |
| `deepseek-ocr2-bf16` | `sabafallah/DeepSeek-OCR-2-GGUF@d08e5af400c64fa8a9b89b04ba373b600b02e05d`
`mmproj-deepseek-ocr-2-bf16.gguf` | 929,037,632 | `b65e8460acc82dd4e8546206c5abd1abed9c5c582223731ce1585689d95f5cdb` | `deepseekocr2` | `sabafallah/DeepSeek-OCR-2-GGUF@d08e5af400c64fa8a9b89b04ba373b600b02e05d`
`deepseek-ocr-2-Q4_K_M.gguf`
1,950,326,688 bytes | `deepseek-ai/DeepSeek-OCR-2@aaa02f3811945a91062062994c5c4a3f4c0af2b0` |
| `dots-ocr-f16` | `ggml-org/dots.ocr-GGUF@2c093a32ca360a396bc6d87d60408636130b9d9b`
`mmproj-dots.ocr-f16.gguf` | 2,526,296,992 | `f462429dbf41729379df4252f84331c9646d80e53f092ccccfd2cb922a3b544e` | `dots_ocr` | `ggml-org/dots.ocr-GGUF@2c093a32ca360a396bc6d87d60408636130b9d9b`
`dots.ocr-Q8_0.gguf`
1,894,530,272 bytes | `dots-studio/dots.ocr@c0111ce6bc07803dbc267932ffef0ae3a51dc951` |
| `dots3note-prev-f16` | `ggml-org/dots3-note-prev-GGUF@e5e7f6692337c0782c2bb4e2395174fd99879448`
`mmproj-dots3-note-prev-F16.gguf` | 15,546,328,640 | `5bfec6cc2e2fa8ffcc70fc89866345640a297e7046e0913fd4899a3ec06f5ace` | `dots3note_v`, `dots3note_a` | `ggml-org/dots3-note-prev-GGUF@e5e7f6692337c0782c2bb4e2395174fd99879448`
`IQ2_S/dots3-note-prev-IQ2_S-00001-of-00003.gguf`
5,940,800 bytes | `dots-studio/dots3-note-prev@1e1e7b0cd37a3a48a6c8d7fa55d5f9d14377006b` |
| `paddleocr-vl-1.6-bf16` | `PaddlePaddle/PaddleOCR-VL-1.6-GGUF@511b09642bb324401f15f97cc23bc67e8f0a291d`
`PaddleOCR-VL-1.6-GGUF-mmproj.gguf` | 881,770,560 | `204d757d7610d9b3faab10d506d69e5b244e32bf765e2bab2d0167e65e0a058a` | `paddleocr` | `PaddlePaddle/PaddleOCR-VL-1.6-GGUF@511b09642bb324401f15f97cc23bc67e8f0a291d`
`PaddleOCR-VL-1.6-GGUF.gguf`
935,769,056 bytes | `PaddlePaddle/PaddleOCR-VL-1.6@c5630abae1d940eafe0697512a0325494b02ab42` |
| `lightonocr-1b-1025-f16` | `noctrex/LightOnOCR-1B-1025-GGUF@fe9d27bebcd975de319b2129a700791e6c9e00ae`
`mmproj-F16.gguf` | 819,312,608 | `af55bc472ee9b5c409b4545033b5d1810b2a9cd799d53922ad93e5069542cd72` | `lightonocr` | `noctrex/LightOnOCR-1B-1025-GGUF@fe9d27bebcd975de319b2129a700791e6c9e00ae`
`LightOnOCR-1B-1025-Q4_K_S.gguf`
470,781,536 bytes | `lightonai/LightOnOCR-1B-1025@7e3e7b0cb83e237e7d237af5a583a002ea632547` |
| `youtu-vl-4b-bf16` | `tencent/Youtu-VL-4B-Instruct-GGUF@1b7e295135d85a169d93823aa4215faf2c427092`
`mmproj-Youtu-VL-4b-Instruct-BF16.gguf` | 893,397,344 | `1bcb2b7a99687be9a47e9bf27ee96237b73fe94b6fda938b17b18e7d4f92f9f2` | `youtuvl` | `tencent/Youtu-VL-4B-Instruct-GGUF@1b7e295135d85a169d93823aa4215faf2c427092`
`Youtu-VL-4B-Instruct-Q8_0.gguf`
5,211,323,488 bytes | `tencent/Youtu-VL-4B-Instruct@8d30a0e49662a1d628a472b12df264dbcd768753` |
| `granite4-vision-4.1-f16` | `ibm-granite/granite-vision-4.1-4b-GGUF@b1fa14294b0f5cac04c43076d1c4574091abf117`
`mmproj-model-f16.gguf` | 1,162,347,936 | `573dd2579f6043649299f0b2225000a5691d92f320aabe909fb4c6e75450cad2` | `granite4_vision` | `ibm-granite/granite-vision-4.1-4b-GGUF@b1fa14294b0f5cac04c43076d1c4574091abf117`
`granite-vision-4.1-4b-Q4_K_M.gguf`
2,099,510,240 bytes | `ibm-granite/granite-vision-4.1-4b@37d591f06319e8f1638b5adcf58bdf50e0f84f7a` |
The five generic projector evidence pairs total **10,672,691,328 bytes** (sidecars plus paired text GGUFs), below the 16 GiB evidence budget. Runtime remains deferred; four routes have independent nonzero-weight graph parity, while MiniCPM resampler remains component-only and graph-deferred.
Pinned source proofs cover graph semantics that cannot be inferred from tensor names,
including conversion-time permutations, co-resident modality roles, and processor boundaries.
| Evidence ID | Immutable sources | Finding |
|---|---|---|
| `glma-converter-checkpoint-drift` | `ggml-org/llama.cpp@86632248188c106d749fad34a1dcd237c95863d4` `tools/mtmd/models/whisper-enc.cpp`
`ggml-org/llama.cpp@86632248188c106d749fad34a1dcd237c95863d4` `conversion/ultravox.py`
`zai-org/GLM-ASR-Nano-2512@61ba4e0b3309b6656edea3e93e419f7bd5c61957` `config.json` | llama.cpp defines the legacy additive-position GLMA graph, but its converter registers GlmasrModel and requires a top-level merge_factor. The immutable current checkpoint declares GlmAsrForConditionalGeneration, omits merge_factor, and configures a partial-RoPE encoder. Mobius can import a structurally valid legacy sidecar without claiming that this checkpoint can produce one. |
| `qwen3tts-speaker-runtime-boundary` | `ggml-org/llama.cpp@86632248188c106d749fad34a1dcd237c95863d4` `tools/mtmd/models/qwen3tts-spkenc.cpp`
`ggml-org/llama.cpp@86632248188c106d749fad34a1dcd237c95863d4` `conversion/qwen3tts.py`
`ggml-org/llama.cpp@86632248188c106d749fad34a1dcd237c95863d4` `tools/mtmd/mtmd-helper-gen.cpp` | The speaker graph deterministically emits one text-width ECAPA embedding. The downstream helper adds it to the tts_pad embedding inside a hand-built four-section MRoPE prompt, while the converter co-emits a separate stateful qwen3tts_gen namespace. Mobius exports only the exact speaker_encoder role and makes no generated-audio runtime claim. |
| `cogvlm-pinned-graph-source` | `ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `tools/mtmd/models/cogvlm.cpp`
`ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `conversion/cogvlm.py`
`zai-org/cogvlm-chat-hf@e29dc3ba206d524bf8efbfc60d80fc4556ab0e3c` `config.json` | Dedicated fused-QKV CLIP tower, gated projector, and BOI/EOI rows. |
| `exaone4-5-pinned-graph-source` | `ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `tools/mtmd/models/exaone4_5.cpp`
`ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `conversion/exaone.py`
`LGAI-EXAONE/EXAONE-4.5-33B@570aa4b15a4f45ba1133072b45f50198f6e3b4fd` `config.json` | Qwen-style vision tower with GQA and exact windowed spatial merger. |
| `hunyuanvl-pinned-graph-source` | `ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `tools/mtmd/models/hunyuanvl.cpp`
`ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `conversion/hunyuan.py`
`tencent/HunyuanOCR@b7bf72439f11fa076c547edf8777aa85f8e0a027` `config.json` | Dynamic ViT, convolutional perceiver, row-newline, and boundary topology. |
| `janus-pro-pinned-graph-source` | `ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `tools/mtmd/models/siglip.cpp`
`ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `conversion/januspro.py`
`deepseek-community/Janus-Pro-1B@1655280bb75959cc1cb85529a2a8b26e7016072e` `config.json` | Fixed SigLIP tower followed by a two-layer exact-GELU aligner. |
| `kimik25-pinned-graph-source` | `ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `tools/mtmd/models/kimik25.cpp`
`ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `conversion/kimivl.py`
`moonshotai/Kimi-K2.5@4d01dfe0332d63057c186e0b262165819efb6611` `config.json` | Bicubic 3D learned positions, converted 2D RoPE, merge, norm, and MLP. |
| `kimivl-pinned-graph-source` | `ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `tools/mtmd/models/kimivl.cpp`
`ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `conversion/kimivl.py`
`moonshotai/Kimi-VL-A3B-Thinking-2506@aa1730989e7558695b44ee493623e03bd325a994` `config.json` | Learned 2D RoPE ViT followed by patch merge, LayerNorm, and GELU MLP. |
| `lfm2-pinned-graph-source` | `ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `tools/mtmd/models/siglip.cpp`
`LiquidAI/LFM2.5-VL-1.6B@919fde3d022e3f90a4716006f993938ee8c2eb97` `config.json` | Dynamic SigLIP positions, 2x2 pixel unshuffle, LayerNorm, and GELU MLP. |
| `meralion-pinned-graph-source` | `ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `tools/mtmd/models/whisper-enc.cpp`
`ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `conversion/ultravox.py`
`MERaLiON/MERaLiON-2-3B@a03e40e9ae4f45fb3d575ed7f67bd9fd5304920d` `modeling_meralion2.py` | Whisper encoder, stack-15, stacked LayerNorm, and four-linear gated adapter. |
| `mimovl-pinned-graph-source` | `ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `tools/mtmd/models/mimovl.cpp`
`XiaomiMiMo/MiMo-V2.5@63651580ca774f8504f676040460aed3e1244ac1` `config.json` | GQA vision tower with row/column windows, sinks, and F32 down projection. |
| `minicpmv4-6-pinned-graph-source` | `ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `tools/mtmd/models/minicpmv.cpp`
`openbmb/MiniCPM-V-4.6@36f34a661a4bd35d0dc2294cb044d2584646c7d3` `config.json` | Bucketed positions, inserted local-attention merger, and final merger. |
| `minimax-m3-pinned-graph-source` | `ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `tools/mtmd/models/minimax-m3.cpp`
`ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `conversion/minimax.py`
`MiniMaxAI/MiniMax-M3@f0e1c1e04d40177e4673a22097036854f536e9c0` `config.json` | Partial two-axis RoPE ViT with two distinct spatial MLP mergers. |
| `nemotron-v2-vl-pinned-graph-source` | `ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `tools/mtmd/models/nemotron-v2-vl.cpp`
`ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `conversion/nemotron.py`
`nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16@ca9543b126e8bf3176916d3d305ccc415f89fd4d` `config.json` | RADIO registers, fixed positions, patch merge, RMSNorm, and ReLU-squared MLP. |
| `step3vl-pinned-graph-source` | `ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `tools/mtmd/models/step3vl.cpp`
`ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `conversion/step3.py`
`stepfun-ai/Step3-VL-10B@5026053b0c2f5dfaa08fc2d149384162c3c8bca1` `config.json` | Absolute plus axial positions and two convolutional downsamplers. |
| `yasa2-pinned-graph-source` | `ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6` `tools/mtmd/models/yasa2.cpp`
`RekaAI/reka-edge-2603@492c81c225fbf5f3263a8245b00827721b119a13` `convert_reka_vlm_to_gguf.py`
`RekaAI/reka-edge-2603@492c81c225fbf5f3263a8245b00827721b119a13` `config.json` | ConvNeXtV2, pre-pool positions, fixed 8x8 pooling, and GELU MLP. |
These additional immutable files prove artifact availability only. Their routes remain
governed by the capability matrix until tensor mapping and component parity are established.
| Candidate route | Immutable available sidecar | Bytes | SHA-256 |
|---|---|---:|---|
| `lfm2` | `LiquidAI/LFM2-VL-1.6B-GGUF@6121de267003bb4d4f325fe10abdc735aee06747`
`mmproj-LFM2-VL-1.6B-F16.gguf` | 830,339,008 | `b637bfa6060be2bc7503ec23ba48b407843d08c2ca83f52be206ea8563ccbae2` |
| `lfm2` | `LiquidAI/LFM2-VL-1.6B-GGUF@6121de267003bb4d4f325fe10abdc735aee06747`
`mmproj-LFM2-VL-1.6B-Q8_0.gguf` | 564,115,648 | `65ec437db88d65fff93f472d00c145e09880769ac67fedff5cd1c0f8d8301d87` |
| `pixtral` | `ggml-org/pixtral-12b-GGUF@cba1ea4420bc2b4f15f50fdec59e30769880a63c`
`mmproj-pixtral-12b-Q8_0.gguf` | 463,091,616 | `5504fe00067629053e6f99abac05f628c653a50394f4929bcc185bc80a10daf4` |
| Projector string | Modality | Graph role / route | Paired text architecture | Metadata/tensor/graph/runtime | Exactness/evidence |
|---|---|---|---|---|---|
| `adapter` | vision | vision_encoder via `generic_projector` | `chatglm` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`glm-edge-v-2b-adapter-f16` |
| `cogvlm` | vision | vision_encoder via `remaining_vision_projector` | `cogvlm` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`cogvlm-chat-v1.1-f16-header`; source evidence=`cogvlm-pinned-graph-source` |
| `deepseekocr` | vision | vision_encoder via `ocr_projector` | `deepseek2-ocr` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`deepseek-ocr-bf16` |
| `deepseekocr2` | vision | vision_encoder via `ocr_projector` | `deepseek2-ocr` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`deepseek-ocr2-bf16` |
| `dots3note_a` | audio | vision_encoder, audio_encoder via `ocr_projector` | `dots3note` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`dots3note-prev-f16` |
| `dots3note_v` | vision | vision_encoder, audio_encoder via `ocr_projector` | `dots3note` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`dots3note-prev-f16` |
| `dots_ocr` | vision | vision_encoder via `ocr_projector` | `qwen2` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`dots-ocr-f16` |
| `exaone4_5` | vision | vision_encoder via `remaining_vision_projector` | `exaone4` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`exaone4-5-33b-f16-header`; source evidence=`exaone4-5-pinned-graph-source` |
| `gemma3` | vision | vision_encoder via `gemma3` | `gemma3` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`gemma3-4b-f16` |
| `gemma3na` | audio | audio_encoder via `core_vlm_projector` | `gemma3n` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`gemma3n-e4b-f16` |
| `gemma3nv` | vision | vision_encoder via `core_vlm_projector` | `gemma3n` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`gemma3n-e4b-f16` |
| `gemma4a` | audio | audio_encoder via `core_vlm_projector` | `gemma4` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`gemma4-e2b-f16` |
| `gemma4ua` | audio | audio_encoder via `core_vlm_projector` | `gemma4` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`gemma4-unified-12b-f16` |
| `gemma4uv` | vision | vision_encoder via `core_vlm_projector` | `gemma4` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`gemma4-unified-12b-f16` |
| `gemma4v` | vision | vision_encoder via `gemma4` | `gemma4` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`gemma4-e2b-f16` |
| `glm4v` | vision | vision_encoder via `qwen_glm_projector` | `glm4`, `glm4moe` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`glm4v-projector-f16` |
| `glma` | audio | audio_encoder via `qwen_glm_projector` | `llama` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | source evidence=`glma-converter-checkpoint-drift` |
| `granite4_vision` | vision | vision_encoder via `ocr_projector` | `granite` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`granite4-vision-4.1-f16` |
| `granite_speech` | audio | audio_encoder via `audio_projector` | `granite` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`granite-speech-4.1-2b-f16` |
| `hunyuanvl` | vision | vision_encoder via `remaining_vision_projector` | `hunyuan_vl` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`hunyuanocr-bf16-header`; source evidence=`hunyuanvl-pinned-graph-source` |
| `idefics3` | vision | vision_encoder via `core_vlm_projector` | `llama` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`smolvlm-256m-idefics3-f16` |
| `internvl` | vision | vision_encoder via `core_vlm_projector` | `qwen2` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`internvl25-1b-f16` |
| `janus_pro` | vision | vision_encoder via `remaining_vision_projector` | `llama` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`janus-pro-1b-f16-header`; source evidence=`janus-pro-pinned-graph-source` |
| `kimik25` | vision | vision_encoder via `remaining_vision_projector` | `deepseek2` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`kimi-k2-5-f16-header`; source evidence=`kimik25-pinned-graph-source` |
| `kimivl` | vision | vision_encoder via `remaining_vision_projector` | `deepseek2` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`kimi-vl-a3b-f16-header`; source evidence=`kimivl-pinned-graph-source` |
| `ldp` | vision | vision_encoder via `generic_projector` | `llama` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`mobilevlm-1.7b-ldp-f16` |
| `ldpv2` | vision | vision_encoder via `generic_projector` | `llama` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`mobilevlm-v2-1.7b-ldpv2-f16` |
| `lfm2` | vision | vision_encoder via `remaining_vision_projector` | `lfm2` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`lfm2-vl-1-6b-f16-header`; source evidence=`lfm2-pinned-graph-source` |
| `lfm2a` | audio | audio_encoder via `audio_projector` | `lfm2` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`lfm2.5-audio-1.5b-f16` |
| `lightonocr` | vision | vision_encoder via `ocr_projector` | `qwen3` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`lightonocr-1b-1025-f16` |
| `llama4` | vision | vision_encoder via `core_vlm_projector` | `llama4` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`llama4-scout-f16` |
| `meralion` | audio | audio_encoder via `audio_projector` | `gemma2` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | source evidence=`meralion-pinned-graph-source` |
| `mimo_audio` | audio | audio_encoder via `audio_projector` | `mimo2` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | source evidence=`mimo-v2.5-audio-source` |
| `mimovl` | vision | vision_encoder via `remaining_vision_projector` | `mimo2` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`mimo-v2-5-f16-header`; source evidence=`mimovl-pinned-graph-source` |
| `minicpmv4_6` | vision | vision_encoder via `remaining_vision_projector` | `qwen35` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`minicpm-v4-6-bf16-header`; source evidence=`minicpmv4-6-pinned-graph-source` |
| `minimax_m3` | vision | vision_encoder via `remaining_vision_projector` | `minimax-m3` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | source evidence=`minimax-m3-pinned-graph-source` |
| `mlp` | vision | vision_encoder via `generic_projector` | `llama` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`llava-llama3-8b-mlp-f16` |
| `muse-glimmer` | vision | vision_encoder via `muse_glimmer` | `muse-glimmer` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`muse-glimmer-30b-bf16` |
| `musicflamingo` | audio | audio_encoder via `audio_projector` | `qwen2` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`music-flamingo-bf16` |
| `nemotron_v2_vl` | vision | vision_encoder via `remaining_vision_projector` | `nemotron_h` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`nemotron-nano-v2-vl-bf16-header`; source evidence=`nemotron-v2-vl-pinned-graph-source` |
| `paddleocr` | vision | vision_encoder via `ocr_projector` | `paddleocr` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`paddleocr-vl-1.6-bf16` |
| `parakeet` | audio | audio_encoder via `audio_projector` | `nemotron_h` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | source evidence=`nemotron-v2-parakeet-source` |
| `phi4` | vision | — | — | metadata=deferred; tensor_map=deferred; graph=deferred; runtime=deferred | CONFIG_DEFERRED — Exact configuration ownership is not implemented. |
| `pixtral` | vision | vision_encoder via `core_vlm_projector` | `llama` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`pixtral-12b-f16` |
| `pockettts_gen` | gen.audio | — | — | metadata=rejected; tensor_map=rejected; graph=rejected; runtime=rejected | CONFIG_REJECTED — The serialized architecture contract is deliberately refused. |
| `pockettts_spkenc` | audio | speaker_encoder via `audio_projector` | `pockettts` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | source evidence=`pockettts-speaker-source` |
| `qwen2.5o` | audio, vision | vision_encoder, audio_encoder via `qwen_glm_projector` | `qwen2vl` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`qwen25-omni-projector-f16` |
| `qwen2.5vl_merger` | vision | vision_encoder via `qwen_vl` | `qwen2vl` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`qwen25-vl-3b-f16` |
| `qwen2a` | audio | audio_encoder via `qwen_glm_projector` | `qwen2` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`qwen2-audio-projector-f16` |
| `qwen2vl_merger` | vision | vision_encoder via `qwen_vl` | `qwen2vl` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`qwen2-vl-2b-f16` |
| `qwen3a` | audio | audio_encoder via `qwen_glm_projector` | `qwen3vl`, `qwen3vlmoe` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`qwen3-audio-projector-bf16` |
| `qwen3tts_gen` | gen.audio | — | — | metadata=rejected; tensor_map=rejected; graph=rejected; runtime=rejected | CONFIG_REJECTED — The serialized architecture contract is deliberately refused. |
| `qwen3tts_spkenc` | audio | speaker_encoder via `qwen_glm_projector` | `qwen3tts` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | source evidence=`qwen3tts-speaker-runtime-boundary` |
| `qwen3vl_merger` | vision | vision_encoder via `qwen_glm_projector` | `qwen35`, `qwen35moe`, `qwen3vl`, `qwen3vlmoe` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`qwen3-vl-projector-f16` |
| `resampler` | vision | vision_encoder via `generic_projector` | `minicpm` | metadata=supported; tensor_map=supported; graph=deferred; runtime=deferred | artifact pins=`minicpm-v2-resampler-f16` |
| `step3vl` | vision | vision_encoder via `remaining_vision_projector` | `qwen3` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`step3-vl-10b-f16-header`; source evidence=`step3vl-pinned-graph-source` |
| `ultravox` | audio | audio_encoder via `audio_projector` | `llama` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`ultravox-v0.5-f16` |
| `voxtral` | audio | audio_encoder via `audio_projector` | `llama` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | source evidence=`voxtral-mini-3b-source-and-q8-availability` |
| `yasa2` | vision | vision_encoder via `remaining_vision_projector` | `llama` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`yasa2-reka-edge-f16-header`; source evidence=`yasa2-pinned-graph-source` |
| `youtuvl` | vision | vision_encoder via `ocr_projector` | `deepseek2` | metadata=supported; tensor_map=supported; graph=supported; runtime=deferred | artifact pins=`youtu-vl-4b-bf16` |
## Tokenizer pre-types
The pre-type is never sufficient evidence by itself. The generated census preserves all aliases
and gives every route an exact evidence ID or concrete compiled-semantics blocker.
| Exact identifier | Semantic group / pre-type | Default policy | Current status | Evidence / blocker |
|---|---|---|---|---|
| `a.x-4.0` | `gpt-2` / `GPT2` | `deferred` | `validated-pinned-source` | `gpt2-q4-tokenizer`; `llama-vocab.cpp:L2194` |
| `afmoe` | `afmoe` / `AFMOE` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `bailingmoe` | `bailingmoe` / `BAILINGMOE` | `deferred` | `deferred-pinned-artifact-mismatch` | `pinned-candidate-source-semantic-mismatch` (`llada-moe-iq1-s-tokenizer-semantic-blocker`); `mradermacher/LLaDA-MoE-7B-A1B-Instruct-i1-GGUF@2ec29fbe69f07f382a864f93b40c4eecb45e6a0a` / `LLaDA-MoE-7B-A1B-Instruct.i1-IQ1_S.gguf` vs `inclusionAI/LLaDA-MoE-7B-A1B-Instruct@67004f662901b09f729994d4b3c04201283941ba`: fail-closed: the exact GGUF and official source agree on ordered vocabulary prefix, merges, token types, added/special tokens, deterministic padding, config, and chat template, but the official NFC normalizer encodes decomposed e-acute as token 2900 while pinned BAILINGMOE preserves tokens 68 and 150766; architecture `llada-moe`, declared pre `llada-moe`, effective pre `llada-moe`; `llama-vocab.cpp:L2336` |
| `bailingmoe2` | `bailingmoe` / `BAILINGMOE` | `deferred` | `deferred-pinned-artifact-mismatch` | `pinned-candidate-source-semantic-mismatch` (`llada-moe-iq1-s-tokenizer-semantic-blocker`); `mradermacher/LLaDA-MoE-7B-A1B-Instruct-i1-GGUF@2ec29fbe69f07f382a864f93b40c4eecb45e6a0a` / `LLaDA-MoE-7B-A1B-Instruct.i1-IQ1_S.gguf` vs `inclusionAI/LLaDA-MoE-7B-A1B-Instruct@67004f662901b09f729994d4b3c04201283941ba`: fail-closed: the exact GGUF and official source agree on ordered vocabulary prefix, merges, token types, added/special tokens, deterministic padding, config, and chat template, but the official NFC normalizer encodes decomposed e-acute as token 2900 while pinned BAILINGMOE preserves tokens 68 and 150766; architecture `llada-moe`, declared pre `llada-moe`, effective pre `llada-moe`; `llama-vocab.cpp:L2337` |
| `bloom` | `bloom` / `BLOOM` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `chameleon` | `chameleon` / `CHAMELEON` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `chatglm-bpe` | `glm4` / `CHATGLM4` | `deferred` | `deferred-pinned-artifact-mismatch` | `pinned-candidate-source-semantic-mismatch` (`glm4-7-flash-iq2-xxs-tokenizer-semantic-blocker`); `bartowski/zai-org_GLM-4.7-Flash-GGUF@464d07505b441959737cd04d900f047469614c8d` / `zai-org_GLM-4.7-Flash-IQ2_XXS.gguf` vs `zai-org/GLM-4.7-Flash@a9308079ef95921451a690cd2d16cb572e564642`: fail-closed: ordered vocabulary, merges, deterministic padding, config, and chat template are pinned, but seven source added-token flags disagree with GGUF types; pinned CHATGLM4 splits apostrophe-space-CJK differently and detokenizes token 659 without the official leading space; architecture `deepseek2`, declared pre `glm4`, effective pre `glm4`; `llama-vocab.cpp:L2257` |
| `codeshell` | `codeshell` / `CODESHELL` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `cohere2moe` | `tiny_aya` / `TINY_AYA` | `deferred` | `deferred-pinned-artifact-mismatch` | `pinned-candidate-source-semantic-mismatch` (`north-mini-code-iq1-s-tokenizer-semantic-blocker`); `mradermacher/North-Mini-Code-1.0-i1-GGUF@94d8eb17eaeb728f907639ee0eff457e3e274667` / `North-Mini-Code-1.0.i1-IQ1_S.gguf` vs `CohereLabs/North-Mini-Code-1.0@d11e61a842617a22dc328552fa5bb86231ee4f37`: fail-closed: ordered vocabulary, merges, deterministic padding, config, and chat template are pinned, but 23 source added-token flags disagree with GGUF types and pinned TINY_AYA splits tab-space differently from the official North tokenizer; architecture `cohere2moe`, declared pre `cohere2moe`, effective pre `cohere2moe`; `llama-vocab.cpp:L2320` |
| `command-r` | `command-r` / `COMMAND_R` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `dbrx` | `dbrx` / `DBRX` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `deepseek-coder` | `deepseek-coder` / `DEEPSEEK_CODER` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `deepseek-llm` | `deepseek-llm` / `DEEPSEEK_LLM` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `deepseek-r1-qwen` | `qwen2` / `QWEN2` | `deferred` | `validated-pinned-source` | `qwen2.5-0.5b-instruct-q8-tokenizer`; `llama-vocab.cpp:L2230` |
| `deepseek-v3` | `deepseek-v3` / `DEEPSEEK3_LLM` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `default` | `default` / `DEFAULT` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `exaone` | `exaone` / `EXAONE` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `exaone-moe` | `exaone-moe` / `EXAONE_MOE` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `exaone4` | `gpt-2` / `GPT2` | `deferred` | `validated-pinned-source` | `gpt2-q4-tokenizer`; `llama-vocab.cpp:L2290` |
| `f2llmv2` | `qwen2` / `QWEN2` | `deferred` | `validated-pinned-source` | `qwen2.5-0.5b-instruct-q8-tokenizer`; `llama-vocab.cpp:L2232` |
| `falcon` | `falcon` / `FALCON` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `falcon-h1` | `llama3` / `LLAMA3` | `deferred` | `validated-pinned-source` | `lfm2-350m-f16-tokenizer`; `llama-vocab.cpp:L2152` |
| `falcon3` | `llama3` / `LLAMA3` | `deferred` | `validated-pinned-source` | `lfm2-350m-f16-tokenizer`; `llama-vocab.cpp:L2151` |
| `gemma4` | `gemma4` / `GEMMA4` | `deferred` | `validated-pinned-source` | `gemma4-e2b-iq2-native-tokenizer`; `llama-vocab.cpp:L2202` |
| `gigachat` | `gpt-2` / `GPT2` | `deferred` | `validated-pinned-source` | `gpt2-q4-tokenizer`; `llama-vocab.cpp:L2191` |
| `glm4` | `glm4` / `CHATGLM4` | `deferred` | `deferred-pinned-artifact-mismatch` | `pinned-candidate-source-semantic-mismatch` (`glm4-7-flash-iq2-xxs-tokenizer-semantic-blocker`); `bartowski/zai-org_GLM-4.7-Flash-GGUF@464d07505b441959737cd04d900f047469614c8d` / `zai-org_GLM-4.7-Flash-IQ2_XXS.gguf` vs `zai-org/GLM-4.7-Flash@a9308079ef95921451a690cd2d16cb572e564642`: fail-closed: ordered vocabulary, merges, deterministic padding, config, and chat template are pinned, but seven source added-token flags disagree with GGUF types; pinned CHATGLM4 splits apostrophe-space-CJK differently and detokenizes token 659 without the official leading space; architecture `deepseek2`, declared pre `glm4`, effective pre `glm4`; `llama-vocab.cpp:L2256` |
| `gpt-2` | `gpt-2` / `GPT2` | `deferred` | `validated-pinned-source` | `gpt2-q4-tokenizer`; `llama-vocab.cpp:L2187` |
| `gpt-4o` | `gpt-4o` / `GPT4O` | `deferred` | `deferred-pinned-artifact-evidence` | `pinned-candidate-identifier-mismatch`; `mradermacher/oh-dcft-v3.1-gpt-4o-mini-GGUF@41c1d48055e3192a907c0ffc2a886288e9040e33` / `oh-dcft-v3.1-gpt-4o-mini.Q2_K.gguf` vs `Xenova/gpt-4o@7956d98f2a83b2751a98ea7136fdf7fe6cf54e69`: name-only candidate rejected: its complete GGUF header dispatches llama-bpe, not gpt-4o, and has 128256 tokens; the pinned llama.cpp fingerprint source has 200000 vocabulary entries plus 2 added tokens and is tokenizer-only, with no model config, chat template, or embedding rows; `llama-vocab.cpp:L2307` |
| `gpt3-finnish` | `gpt3-finnish` / `GPT3_FINNISH` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `granite-docling` | `granite-docling` / `GRANITE_DOCLING` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `granite-embed-multi-311m` | `gemma4` / `GEMMA4` | `deferred` | `deferred-pinned-artifact-evidence` | `pinned-candidate-effective-pre-mismatch`; `SandLogicTechnologies/granite-embedding-311m-multilingual-r2-GGUF@f142535239859391fbef67aaf886d96500ad9fa8` / `granite-embedding-311M-multilingual-r2_IQ4_XS.gguf` vs `ibm-granite/granite-embedding-311m-multilingual-r2@44399559930365213510b1ee2eb15ded83374f0e`: complete modern-bert artifact rejected for this exact identifier: it declares no tokenizer.ggml.pre, so pinned llama.cpp uses fallback gemma4 rather than granite-embed-multi-311m. Its artifact-scoped fallback reconstruction matches 480 pinned tokenize/detokenize cases. GGUF-only IDs 262145..262151 are type-4 user-defined tokens; fallback gemma4 promotes only ID 262149 (<|tool_response>) to control/EOG. This cannot prove the absent identifier; architecture `modern-bert`, declared pre `absent`, effective pre `gemma4`; `llama-vocab.cpp:L2203` |
| `granite-embed-multi-97m` | `granite-embed-multi-97m` / `GRANITE_EMB_MULTI` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `grok-2` | `grok-2` / `GROK_2` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `hunyuan` | `hunyuan` / `HUNYUAN` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `hunyuan-dense` | `hunyuan-dense` / `HUNYUAN_DENSE` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `jais` | `jais` / `JAIS` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `jais-2` | `jais-2` / `JAIS2` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `jina-de` | `gpt-2` / `GPT2` | `deferred` | `validated-pinned-source` | `gpt2-q4-tokenizer`; `llama-vocab.cpp:L2190` |
| `jina-es` | `gpt-2` / `GPT2` | `deferred` | `validated-pinned-source` | `gpt2-q4-tokenizer`; `llama-vocab.cpp:L2189` |
| `jina-v1-en` | `jina-v1-en` / `GPT2_ADD_SEP` | `deferred` | `deferred-pinned-artifact-evidence` | `pinned-candidate-source-token-mismatch`; `gpustack/jina-reranker-v1-tiny-en-GGUF@34fdafe5a08b64246bcbfdbf0b8a23f818baf8e3` / `jina-reranker-v1-tiny-en-Q2_K.gguf` vs `jinaai/jina-reranker-v1-tiny-en@aca45de6945b5dc6399abcd2a9c55ded5dc9111f`: ordered token id 5 differs: GGUF is empty while the official tokenizer is U+0000; deterministic padding starts only at id 60516; `llama-vocab.cpp:L2212` |
| `jina-v2-code` | `jina-v1-en` / `GPT2_ADD_SEP` | `deferred` | `validated-pinned-source` | `jina-v2-code-q8-tokenizer`; `llama-vocab.cpp:L2213` |
| `jina-v2-de` | `gpt-2` / `GPT2` | `deferred` | `validated-pinned-source` | `gpt2-q4-tokenizer`; `llama-vocab.cpp:L2193` |
| `jina-v2-es` | `gpt-2` / `GPT2` | `deferred` | `validated-pinned-source` | `gpt2-q4-tokenizer`; `llama-vocab.cpp:L2192` |
| `jina-v5-nano` | `llama3` / `LLAMA3` | `deferred` | `validated-pinned-source` | `lfm2-350m-f16-tokenizer`; `llama-vocab.cpp:L2156` |
| `joyai-llm` | `joyai-llm` / `JOYAI_LLM` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `kanana2` | `gpt-4o` / `GPT4O` | `deferred` | `validated-pinned-source` | `kanana2-1.3b-instruct-q8-tokenizer`; `llama-vocab.cpp:L2309` |
| `kimi-k2` | `kimi-k2` / `KIMI_K2` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `kormo` | `qwen2` / `QWEN2` | `deferred` | `validated-pinned-source` | `qwen2.5-0.5b-instruct-q8-tokenizer`; `llama-vocab.cpp:L2231` |
| `laguna` | `laguna` / `LAGUNA` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `lfm2` | `llama3` / `LLAMA3` | `deferred` | `validated-pinned-source` | `lfm2-350m-f16-tokenizer`; `llama-vocab.cpp:L2155` |
| `llada-moe` | `bailingmoe` / `BAILINGMOE` | `deferred` | `deferred-pinned-artifact-mismatch` | `pinned-candidate-source-semantic-mismatch` (`llada-moe-iq1-s-tokenizer-semantic-blocker`); `mradermacher/LLaDA-MoE-7B-A1B-Instruct-i1-GGUF@2ec29fbe69f07f382a864f93b40c4eecb45e6a0a` / `LLaDA-MoE-7B-A1B-Instruct.i1-IQ1_S.gguf` vs `inclusionAI/LLaDA-MoE-7B-A1B-Instruct@67004f662901b09f729994d4b3c04201283941ba`: fail-closed: the exact GGUF and official source agree on ordered vocabulary prefix, merges, token types, added/special tokens, deterministic padding, config, and chat template, but the official NFC normalizer encodes decomposed e-acute as token 2900 while pinned BAILINGMOE preserves tokens 68 and 150766; architecture `llada-moe`, declared pre `llada-moe`, effective pre `llada-moe`; `llama-vocab.cpp:L2338` |
| `llama-bpe` | `llama3` / `LLAMA3` | `deferred` | `validated-pinned-source` | `lfm2-350m-f16-tokenizer`; `llama-vocab.cpp:L2150` |
| `llama-v3` | `llama3` / `LLAMA3` | `deferred` | `validated-pinned-source` | `lfm2-350m-f16-tokenizer`; `llama-vocab.cpp:L2149` |
| `llama3` | `llama3` / `LLAMA3` | `deferred` | `validated-pinned-source` | `lfm2-350m-f16-tokenizer`; `llama-vocab.cpp:L2148` |
| `llama4` | `gpt-4o` / `GPT4O` | `deferred` | `deferred-pinned-artifact-evidence` | `pinned-candidate-incomplete-shard`; `ggml-org/Llama-4-Scout-17B-16E-Instruct-GGUF@42675345da11ade9203a5187595da7b74d4ff2ac` / `Llama-4-Scout-17B-16E-Instruct-Q4_K_M-00002-of-00002.gguf` vs `meta-llama/Llama-4-Scout-17B-16E-Instruct@92f3b1597a195b523d8d9e5700e57e4fbb8f20d3`: the only pinned Q4_K_M file within 16 GiB is shard 2 of 2; its complete header has only split metadata, 145 of 628 tensors, no tokenizer fields, and no embedding tensor, while shard 1 is 49848377344 bytes; `llama-vocab.cpp:L2308` |
| `megrez` | `megrez` / `QWEN2_CLEAN_SPACES` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `mellum` | `gpt-2` / `GPT2` | `deferred` | `validated-pinned-source` | `gpt2-q4-tokenizer`; `llama-vocab.cpp:L2195` |
| `mellum2` | `mellum2` / `MELLUM2` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `midm-2.0` | `llama3` / `LLAMA3` | `deferred` | `validated-pinned-source` | `lfm2-350m-f16-tokenizer`; `llama-vocab.cpp:L2154` |
| `minerva-7b` | `minerva-7b` / `MINERVA` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `minicpm5` | `minicpm5` / `MINICPM5` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `minimax-m2` | `minimax-m2` / `MINIMAX_M2` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `modern-bert` | `gpt-2` / `GPT2` | `deferred` | `validated-pinned-source` | `gpt2-q4-tokenizer`; `llama-vocab.cpp:L2196` |
| `mpt` | `mpt` / `MPT` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `olmo` | `olmo` / `OLMO` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `phi-2` | `gpt-2` / `GPT2` | `deferred` | `validated-pinned-source` | `gpt2-q4-tokenizer`; `llama-vocab.cpp:L2188` |
| `pixtral` | `llama3` / `LLAMA3` | `deferred` | `validated-pinned-source` | `lfm2-350m-f16-tokenizer`; `llama-vocab.cpp:L2153` |
| `poro-chat` | `poro-chat` / `PORO` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `qwen2` | `qwen2` / `QWEN2` | `deferred` | `validated-pinned-source` | `qwen2.5-0.5b-instruct-q8-tokenizer`; `llama-vocab.cpp:L2229` |
| `qwen35` | `qwen35` / `QWEN35` | `deferred` | `validated-pinned-source` | `qwen3.5-0.8b-q4-tokenizer` |
| `refact` | `refact` / `REFACT` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `roberta-bpe` | `jina-v1-en` / `GPT2_ADD_SEP` | `deferred` | `validated-pinned-source` | `roberta-bpe-q2-tokenizer`; `llama-vocab.cpp:L2214` |
| `sarvam-moe` | `sarvam-moe` / `SARVAM_MOE` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `seed-coder` | `seed-coder` / `SEED_CODER` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `smaug-bpe` | `smaug-bpe` / `SMAUG` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `smollm` | `smollm` / `SMOLLM` | `deferred` | `validated-pinned-source` | `smollm-135m-f16-tokenizer` |
| `solar-open` | `solar-open` / `SOLAR_OPEN` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `stablelm2` | `stablelm2` / `STABLELM2` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `starcoder` | `starcoder` / `STARCODER` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `superbpe` | `superbpe` / `SUPERBPE` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `talkie` | `gpt-4o` / `GPT4O` | `deferred` | `validated-pinned-source` | `talkie-13b-q4-native-tokenizer`; `llama-vocab.cpp:L2310` |
| `tekken` | `tekken` / `TEKKEN` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `tiny_aya` | `tiny_aya` / `TINY_AYA` | `deferred` | `deferred-pinned-artifact-mismatch` | `pinned-candidate-source-semantic-mismatch` (`north-mini-code-iq1-s-tokenizer-semantic-blocker`); `mradermacher/North-Mini-Code-1.0-i1-GGUF@94d8eb17eaeb728f907639ee0eff457e3e274667` / `North-Mini-Code-1.0.i1-IQ1_S.gguf` vs `CohereLabs/North-Mini-Code-1.0@d11e61a842617a22dc328552fa5bb86231ee4f37`: fail-closed: ordered vocabulary, merges, deterministic padding, config, and chat template are pinned, but 23 source added-token flags disagree with GGUF types and pinned TINY_AYA splits tab-space differently from the official North tokenizer; architecture `cohere2moe`, declared pre `cohere2moe`, effective pre `cohere2moe`; `llama-vocab.cpp:L2319` |
| `trillion` | `trillion` / `TRILLION` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `viking` | `viking` / `VIKING` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `whitespace` | `whitespace` / `WHITESPACE` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
| `youtu` | `youtu` / `YOUTU` | `deferred` | `deferred-compiled-semantics` | `compiled-llama.cpp-semantic-dependency` |
## Validation boundary
Normal tests are deterministic and network-free; committed registry records contain compact
immutable identities and semantic hashes. Real-artifact qualification is performed serially
with pinned revisions, full SHA-256 verification, at least twice the artifact size free, and
independent runtime evidence where runtime support is claimed.