qwen3_5_moe_vl

Model type	qwen3_5_moe_vl
Class	Qwen35MoEVL3ModelCausalLMModel
Task	hybrid-qwen-vl
Source	models/qwen35.py

Description

Qwen3.5-MoE-VL vision-language model (3-model split for ORT GenAI).

Builds three separate ONNX models:

• decoder: text decoder over hybrid linear/full attention with MoE FFN; consumes inputs_embeds so the embedding model can splice in vision features.

• vision_encoder: shared Qwen3-VL ViT (identical to the dense Qwen3.5-VL counterpart).

• embedding: token embedding + image-feature fusion.

HuggingFace class: Qwen3_5MoeForConditionalGeneration. The HF model_type string is qwen3_5_moe for both the text-only checkpoints and these VL checkpoints; the registry dispatches to this class when the HF config carries a vision_config sub-object and to

class:

Qwen35MoECausalLMModel otherwise.

Usage

mobius build --model <MODEL_ID> output_dir/

from mobius import build

model = build("<MODEL_ID>")