qwen3_vl_single¶


Model type	`qwen3_vl_single`
Class	`Qwen3VLCausalLMModel`
Task	`qwen3-vl-vision-language`
Source	`models/qwen_vl.py`

Description¶

Qwen3-VL vision-language model with packed-attention vision encoder.

Combines a Qwen3-VL ViT vision encoder with a Qwen3-VL text decoder. The vision encoder processes packed image/video patches through Conv3d embedding, rotary-embedded transformer blocks with packed attention, spatial merge, and DeepStack intermediate feature extraction.

The text decoder uses interleaved MRoPE for 3D positional encoding (temporal, height, width) and injects DeepStack vision features into early decoder layers at image token positions.

Weight names match HuggingFace convention: visual.* for vision encoder, language_model.* for text decoder.

Usage¶

mobius build --model <MODEL_ID> output_dir/

from mobius import build

model = build("<MODEL_ID>")