Supported Models¶

mobius supports 298 registered model types.

Text Generation¶

Standard autoregressive language models (CausalLM).

`model_type`	Class	Task
DFlashDraftModel	`DFlashDraftModel`	`dflash-draft`
Eagle3DraftModel	`Eagle3DraftModel`	`eagle3-draft`
Eagle3LlamaForCausalLM	`Eagle3DraftModel`	`eagle3-draft`
Eagle3Speculator	`Eagle3DraftModel`	`eagle3-draft`
Gemma4AssistantForCausalLM	`Gemma4AssistantCausalLMModel`	`gemma4-assistant`
Gemma4UnifiedAssistantForCausalLM	`Gemma4AssistantCausalLMModel`	`gemma4-assistant`
LlamaForCausalLMEagle3	`Eagle3DraftModel`	`eagle3-draft`
Qwen35MtpModel	`Qwen35MtpModel`	`qwen35-mtp`
apertus	`ApertusCausalLMModel`	`text-generation`
arcee	`ArceeCausalLMModel`	`text-generation`
baichuan	`CausalLMModel`	`text-generation`
bloom	`BloomCausalLMModel`	`text-generation`
chatglm	`ChatGLMCausalLMModel`	`text-generation`
code_llama	`CausalLMModel`	`text-generation`
codegen	`CodeGenCausalLMModel`	`text-generation`
codegen2	`CausalLMModel`	`text-generation`
cohere	`CohereCausalLMModel`	`text-generation`
cohere2	`CohereCausalLMModel`	`text-generation`
command_r	`CausalLMModel`	`text-generation`
csm	`CausalLMModel`	`text-generation`
diffllama	`DiffLlamaCausalLMModel`	`text-generation`
doge	`DogeCausalLMModel`	`text-generation`
ernie4_5	`ErnieCausalLMModel`	`text-generation`
evolla	`CausalLMModel`	`text-generation`
exaone	`CausalLMModel`	`text-generation`
exaone4	`ExaOne4CausalLMModel`	`text-generation`
falcon	`FalconCausalLMModel`	`text-generation`
falcon_h1	`FalconCausalLMModel`	`text-generation`
falcon_mamba	`MambaCausalLMModel`	`ssm-text-generation`
gemma	`GemmaCausalLMModel`	`text-generation`
gemma2	`Gemma2CausalLMModel`	`text-generation`
gemma3_text	`Gemma3CausalLMModel`	`text-generation`
gemma3n	`Gemma3nCausalLMModel`	`text-generation`
gemma3n_text	`Gemma3nCausalLMModel`	`text-generation`
gemma4_assistant	`Gemma4AssistantCausalLMModel`	`gemma4-assistant`
gemma4_unified_assistant	`Gemma4AssistantCausalLMModel`	`gemma4-assistant`
glm	`GlmCausalLMModel`	`text-generation`
glm4	`Glm4CausalLMModel`	`text-generation`
glm4v_text	`Glm4CausalLMModel`	`text-generation`
gpt_neox	`GPTNeoXCausalLMModel`	`text-generation`
gpt_neox_japanese	`GPTNeoXJapaneseCausalLMModel`	`text-generation`
gptj	`GPTJCausalLMModel`	`text-generation`
granite	`GraniteCausalLMModel`	`text-generation`
helium	`CausalLMModel`	`text-generation`
hunyuan_v1_dense	`HunYuanV1DenseCausalLMModel`	`text-generation`
internlm2	`InternLM2CausalLMModel`	`text-generation`
llada	`LLaDAModel`	`masked-diffusion`
llama	`CausalLMModel`	`text-generation`
llama4_text	`Llama4CausalLMModel`	`text-generation`
mamba	`MambaCausalLMModel`	`ssm-text-generation`
mamba2	`Mamba2CausalLMModel`	`ssm2-text-generation`
minicpm	`CausalLMModel`	`text-generation`
minicpm3	`CausalLMModel`	`text-generation`
minimax	`MiniMaxCausalLMModel`	`hybrid-text-generation`
ministral	`CausalLMModel`	`text-generation`
ministral3	`CausalLMModel`	`text-generation`
mistral	`CausalLMModel`	`text-generation`
modernbert-decoder	`ModernBertDecoderModel`	`text-generation`
mpt	`MPTCausalLMModel`	`text-generation`
nanochat	`NanoChatCausalLMModel`	`text-generation`
nemotron	`NemotronCausalLMModel`	`text-generation`
olmo	`OLMoCausalLMModel`	`text-generation`
olmo2	`OLMo2CausalLMModel`	`text-generation`
olmo3	`OLMo2CausalLMModel`	`text-generation`
open-llama	`CausalLMModel`	`text-generation`
openelm	`CausalLMModel`	`text-generation`
persimmon	`PersimmonCausalLMModel`	`text-generation`
phi	`PhiCausalLMModel`	`text-generation`
phi3	`Phi3CausalLMModel`	`text-generation`
phi3small	`Phi3SmallCausalLMModel`	`text-generation`
qwen	`QwenCausalLMModel`	`text-generation`
qwen2	`CausalLMModel`	`text-generation`
qwen2_5_vl_text	`Qwen25VLTextModel`	`text-generation`
qwen2_vl_text	`Qwen25VLTextModel`	`text-generation`
qwen3	`Qwen3CausalLMModel`	`text-generation`
qwen3_5_text	`Qwen35CausalLMModel`	`hybrid-text-generation`
qwen3_5_vl_text	`Qwen35VLTextModel`	`hybrid-text-generation`
qwen3_tts_tokenizer_12hz	`Qwen3TTSTokenizerV2Model`	`text-generation`
qwen3_vl_text	`Qwen3VLTextModel`	`text-generation`
seed_oss	`CausalLMModel`	`text-generation`
shieldgemma2	`Gemma2CausalLMModel`	`text-generation`
smollm3	`SmolLM3CausalLMModel`	`text-generation`
solar_open	`CausalLMModel`	`text-generation`
stablelm	`LayerNormCausalLMModel`	`text-generation`
starcoder2	`StarCoder2CausalLMModel`	`text-generation`
yi	`CausalLMModel`	`text-generation`
zamba	`CausalLMModel`	`text-generation`

Mixture of Experts¶

Models that route tokens to a subset of expert MLPs.

`model_type`	Class	Task
arctic	`MoECausalLMModel`	`text-generation`
dbrx	`MoECausalLMModel`	`text-generation`
deepseek_v2	`DeepSeekV3CausalLMModel`	`text-generation`
deepseek_v2_moe	`DeepSeekV3CausalLMModel`	`text-generation`
deepseek_v3	`DeepSeekV3CausalLMModel`	`text-generation`
deepseek_v4	`DeepSeekV4CausalLMModel`	`deepseek-v4`
dots1	`DeepSeekV3CausalLMModel`	`text-generation`
ernie4_5_moe	`Ernie45MoECausalLMModel`	`text-generation`
flex_olmo	`MoECausalLMModel`	`text-generation`
glm4_moe	`Glm4MoECausalLMModel`	`text-generation`
glm4v_moe_text	`Glm4MoECausalLMModel`	`text-generation`
gpt_oss	`GPTOSSCausalLMModel`	`text-generation`
granitemoe	`GraniteMoECausalLMModel`	`text-generation`
granitemoeshared	`GraniteMoECausalLMModel`	`text-generation`
hunyuan_v1_moe	`HunYuanMoEV1CausalLMModel`	`text-generation`
jetmoe	`JetMoeCausalLMModel`	`text-generation`
longcat_flash	`LongcatFlashCausalLMModel`	`text-generation`
mixtral	`MoECausalLMModel`	`text-generation`
olmoe	`MoECausalLMModel`	`text-generation`
phimoe	`Phi3MoECausalLMModel`	`text-generation`
qwen2_moe	`Qwen2MoECausalLMModel`	`text-generation`
qwen3_5_moe	`Qwen35MoECausalLMModel`	`hybrid-text-generation`
qwen3_moe	`MoECausalLMModel`	`text-generation`
qwen3_next	`Qwen3NextCausalLMModel`	`hybrid-text-generation`
qwen3_omni_moe	`MoECausalLMModel`	`text-generation`
qwen3_vl_moe	`MoECausalLMModel`	`text-generation`
youtu	`DeepSeekV3CausalLMModel`	`text-generation`

Multimodal¶

Models that process images, audio, or other modalities alongside text.

`model_type`	Class	Task
aya_vision	`LLaVAModel`	`vision-language`
blip-2	`Blip2Model`	`vision-language`
chameleon	`LLaVAModel`	`vision-language`
cohere2_vision	`LLaVAModel`	`vision-language`
deepseek_vl	`LLaVAModel`	`vision-language`
deepseek_vl_hybrid	`LLaVAModel`	`vision-language`
deepseek_vl_v2	`DeepSeekOCR2CausalLMModel`	`vision-language`
florence2	`LLaVAModel`	`vision-language`
fuyu	`LLaVAModel`	`vision-language`
gemma3	`Gemma3MultiModalModel`	`vision-language`
gemma4	`Gemma4Model`	`gemma4`
gemma4_unified	`Gemma4UnifiedModel`	`gemma4-unified`
glm4v	`LLaVAModel`	`vision-language`
glm4v_moe	`LLaVAModel`	`vision-language`
got_ocr2	`LLaVAModel`	`vision-language`
hunyuan_vl_mot	`HunYuanVLMoTModel`	`hunyuan-vl-mot`
idefics2	`LLaVAModel`	`vision-language`
idefics3	`LLaVAModel`	`vision-language`
instructblip	`LLaVAModel`	`vision-language`
instructblipvideo	`LLaVAModel`	`vision-language`
internvl	`InternVL2Model`	`vision-language`
internvl2	`InternVL2Model`	`vision-language`
internvl_chat	`InternVL2Model`	`vision-language`
janus	`LLaVAModel`	`vision-language`
llava	`LLaVAModel`	`vision-language`
llava_next	`LLaVAModel`	`vision-language`
llava_next_video	`LLaVAModel`	`vision-language`
llava_onevision	`LLaVAModel`	`vision-language`
mistral3	`LLaVAModel`	`vision-language`
mllama	`MllamaCausalLMModel`	`mllama-vision-language`
molmo	`LLaVAModel`	`vision-language`
ovis2	`LLaVAModel`	`vision-language`
paligemma	`LLaVAModel`	`vision-language`
phi3_v	`Phi3VModel`	`vision-language`
phi4-siglip	`Phi4SigLIPModel`	`vision-language`
phi4_multimodal	`Phi4MMMultiModalModel`	`phi4mm-multimodal`
phi4mm	`Phi4MMMultiModalModel`	`phi4mm-multimodal`
pixtral	`LLaVAModel`	`vision-language`
qwen2_5_vl	`Qwen25VLCausalLMModel`	`qwen-vl`
qwen2_vl	`Qwen2VLCausalLMModel`	`qwen-vl`
qwen3_5	`Qwen35VL3ModelCausalLMModel`	`hybrid-qwen-vl`
qwen3_5_moe_vl	`Qwen35MoEVL3ModelCausalLMModel`	`hybrid-qwen-vl`
qwen3_5_vl	`Qwen35VL3ModelCausalLMModel`	`hybrid-qwen-vl`
qwen3_vl	`Qwen3VL3ModelCausalLMModel`	`qwen-vl`
qwen3_vl_single	`Qwen3VLCausalLMModel`	`qwen3-vl-vision-language`
smolvlm	`LLaVAModel`	`vision-language`
video_llava	`LLaVAModel`	`vision-language`
vipllava	`LLaVAModel`	`vision-language`

Speech-to-Text¶

Encoder-decoder models for speech recognition.

`model_type`	Class	Task
fastconformer_rnnt	`EncDecRNNTModel`	`fastconformer-rnnt`
fun_asr	`FunASRForConditionalGeneration`	`fun-asr-speech-language`
mms	`Wav2Vec2ForCTCModel`	`ctc-asr`
qwen3_asr	`Qwen3ASRForConditionalGeneration`	`speech-language`
qwen3_forced_aligner	`Qwen3ASRForConditionalGeneration`	`speech-language`
sensevoice_small	`SenseVoiceSmallModel`	`audio-ctc`
whisper	`WhisperForConditionalGeneration`	`speech-to-text`

Audio¶

Audio encoder models for feature extraction (Wav2Vec2, HuBERT, WavLM).

`model_type`	Class	Task
data2vec-audio	`Wav2Vec2Model`	`audio-feature-extraction`
hubert	`Wav2Vec2Model`	`audio-feature-extraction`
mctct	`Wav2Vec2Model`	`audio-feature-extraction`
musicgen	`Wav2Vec2Model`	`audio-feature-extraction`
qwen3_tts	`Qwen3TTSForConditionalGeneration`	`tts`
seamless_m4t	`Wav2Vec2Model`	`audio-feature-extraction`
seamless_m4t_v2	`Wav2Vec2Model`	`audio-feature-extraction`
sew	`Wav2Vec2Model`	`audio-feature-extraction`
sew-d	`Wav2Vec2Model`	`audio-feature-extraction`
speecht5	`Wav2Vec2Model`	`audio-feature-extraction`
unispeech	`Wav2Vec2Model`	`audio-feature-extraction`
unispeech-sat	`Wav2Vec2Model`	`audio-feature-extraction`
voxtral_encoder	`Wav2Vec2Model`	`audio-feature-extraction`
wav2vec2	`Wav2Vec2Model`	`audio-feature-extraction`
wav2vec2-bert	`Wav2Vec2Model`	`audio-feature-extraction`
wav2vec2-conformer	`Wav2Vec2Model`	`audio-feature-extraction`
wavlm	`Wav2Vec2Model`	`audio-feature-extraction`

encoder-only¶

Encoder-only models for embeddings and classification (BERT, RoBERTa).

`model_type`	Class	Task
distilbert	`DistilBertModel`	`feature-extraction`

encoder¶

Encoder-only transformer models (BERT family).

`model_type`	Class	Task
albert	`BertModel`	`feature-extraction`
bert	`BertModel`	`feature-extraction`
bros	`BertModel`	`feature-extraction`
camembert	`BertModel`	`feature-extraction`
data2vec-text	`BertModel`	`feature-extraction`
deberta	`BertModel`	`feature-extraction`
deberta-v2	`BertModel`	`feature-extraction`
electra	`BertModel`	`feature-extraction`
ernie	`BertModel`	`feature-extraction`
ernie_m	`BertModel`	`feature-extraction`
esm	`BertModel`	`feature-extraction`
flaubert	`BertModel`	`feature-extraction`
ibert	`BertModel`	`feature-extraction`
layoutlm	`BertModel`	`feature-extraction`
layoutlmv2	`BertModel`	`feature-extraction`
layoutlmv3	`LayoutLMv3Model`	`feature-extraction`
lilt	`BertModel`	`feature-extraction`
markuplm	`BertModel`	`feature-extraction`
mega	`BertModel`	`feature-extraction`
megatron-bert	`BertModel`	`feature-extraction`
mobilebert	`BertModel`	`feature-extraction`
modernbert	`ModernBertModel`	`feature-extraction`
mpnet	`BertModel`	`feature-extraction`
mra	`BertModel`	`feature-extraction`
nezha	`BertModel`	`feature-extraction`
nystromformer	`BertModel`	`feature-extraction`
qdqbert	`BertModel`	`feature-extraction`
rembert	`BertModel`	`feature-extraction`
roberta	`BertModel`	`feature-extraction`
roberta-prelayernorm	`BertModel`	`feature-extraction`
roc_bert	`BertModel`	`feature-extraction`
roformer	`BertModel`	`feature-extraction`
splinter	`BertModel`	`feature-extraction`
squeezebert	`BertModel`	`feature-extraction`
xlm-roberta	`BertModel`	`feature-extraction`
xlm-roberta-xl	`BertModel`	`feature-extraction`
xlnet	`BertModel`	`feature-extraction`
xmod	`BertModel`	`feature-extraction`
yoso	`BertModel`	`feature-extraction`

encoder-decoder¶

Encoder-decoder sequence-to-sequence models (BART, T5, mBART).

`model_type`	Class	Task
bart	`BartForConditionalGeneration`	`seq2seq`
bigbird_pegasus	`BartForConditionalGeneration`	`seq2seq`
blenderbot	`BartForConditionalGeneration`	`seq2seq`
blenderbot-small	`BartForConditionalGeneration`	`seq2seq`
fsmt	`BartForConditionalGeneration`	`seq2seq`
led	`BartForConditionalGeneration`	`seq2seq`
longt5	`T5ForConditionalGeneration`	`seq2seq`
m2m_100	`BartForConditionalGeneration`	`seq2seq`
marian	`BartForConditionalGeneration`	`seq2seq`
mbart	`BartForConditionalGeneration`	`seq2seq`
mt5	`T5ForConditionalGeneration`	`seq2seq`
mvp	`BartForConditionalGeneration`	`seq2seq`
nllb-moe	`BartForConditionalGeneration`	`seq2seq`
nllb_moe	`BartForConditionalGeneration`	`seq2seq`
pegasus	`BartForConditionalGeneration`	`seq2seq`
pegasus_x	`BartForConditionalGeneration`	`seq2seq`
plbart	`BartForConditionalGeneration`	`seq2seq`
prophetnet	`BartForConditionalGeneration`	`seq2seq`
switch_transformers	`T5ForConditionalGeneration`	`seq2seq`
t5	`T5ForConditionalGeneration`	`seq2seq`
trocr	`TrOCRForConditionalGeneration`	`seq2seq`
umt5	`T5ForConditionalGeneration`	`seq2seq`
xlm-prophetnet	`BartForConditionalGeneration`	`seq2seq`

causal-lm¶

Absolute positional embedding language models (GPT-2 style).

`model_type`	Class	Task
biogpt	`GPT2CausalLMModel`	`text-generation`
ctrl	`CTRLCausalLMModel`	`text-generation`
gpt-sw3	`GPT2CausalLMModel`	`text-generation`
gpt2	`GPT2CausalLMModel`	`text-generation`
gpt_bigcode	`GPT2CausalLMModel`	`text-generation`
gpt_neo	`GPT2CausalLMModel`	`text-generation`
imagegpt	`GPT2CausalLMModel`	`text-generation`
openai-gpt	`GPT2CausalLMModel`	`text-generation`
opt	`OPTCausalLMModel`	`text-generation`
xglm	`GPT2CausalLMModel`	`text-generation`
xlm	`XLMCausalLMModel`	`text-generation`

vision¶

Vision-only models for image classification and feature extraction.

`model_type`	Class	Task
beit	`ViTModel`	`image-classification`
blip	`BlipVisionModel`	`image-classification`
clip_vision_model	`CLIPVisionModel`	`image-classification`
cvt	`ViTModel`	`image-classification`
data2vec-vision	`ViTModel`	`image-classification`
deit	`ViTModel`	`image-classification`
dinov2	`ViTModel`	`image-classification`
dinov2_with_registers	`ViTModel`	`image-classification`
dinov3_vit	`ViTModel`	`image-classification`
hiera	`ViTModel`	`image-classification`
ijepa	`ViTModel`	`image-classification`
mobilevit	`ViTModel`	`image-classification`
mobilevitv2	`ViTModel`	`image-classification`
pvt	`ViTModel`	`image-classification`
pvt_v2	`ViTModel`	`image-classification`
sam2	`Sam2VisionModel`	`image-classification`
siglip	`SigLIPVisionModel`	`image-classification`
siglip2	`SigLIPVisionModel`	`image-classification`
siglip2_vision_model	`SigLIPVisionModel`	`image-classification`
siglip_vision_model	`SigLIPVisionModel`	`image-classification`
swin	`ViTModel`	`image-classification`
swin2sr	`ViTModel`	`image-classification`
swinv2	`ViTModel`	`image-classification`
vit	`ViTModel`	`image-classification`
vit_hybrid	`ViTModel`	`image-classification`
vit_mae	`ViTModel`	`image-classification`
vit_msn	`ViTModel`	`image-classification`

Depth Estimation¶

`model_type`	Class	Task
depth_anything	`DepthAnythingForDepthEstimation`	`image-classification`

Encoder¶

`model_type`	Class	Task
clip_text_model	`CLIPTextModel`	`feature-extraction`

Hybrid SSM+Attention¶

`model_type`	Class	Task
bamba	`BambaCausalLMModel`	`hybrid-text-generation`
granitemoehybrid	`GraniteMoeHybridCausalLMModel`	`hybrid-text-generation`
jamba	`JambaCausalLMModel`	`hybrid-text-generation`
nemotron_h	`NemotronHCausalLMModel`	`hybrid-text-generation`
zamba2	`Zamba2CausalLMModel`	`hybrid-text-generation`

Object Detection¶

`model_type`	Class	Task
yolos	`YolosForObjectDetection`	`object-detection`

Segmentation¶

`model_type`	Class	Task
segformer	`SegformerForSemanticSegmentation`	`image-classification`

Text¶

`model_type`	Class	Task
gemma4_text	`Gemma4CausalLMModel`	`gemma4-text-generation`
gemma4_unified_text	`Gemma4CausalLMModel`	`gemma4-text-generation`