[!NOTE] This is a copy of codecho/Qwen3.5-35B-A3B-text-only at commit 504ce4f, made on 2026-10-30. vLLM CI uses this checkpoint and the original is on a personal account, so we keep a copy here to reduce supply-chain risk.

Qwen3.5-35B-A3B Text-Only

Text-only weights extracted from Qwen/Qwen3.5-35B-A3B (VLM, Mixture-of-Experts) for use with vLLM's Qwen3_5MoeForCausalLM architecture.

What this is

Qwen3.5 MoE models are natively multimodal (VLM). Their HuggingFace checkpoints use Qwen3_5MoeForConditionalGeneration with weights prefixed as model.language_model.*. This repo provides the language model backbone only, with:

• architectures: ["Qwen3_5MoeForCausalLM"]
• model_type: "qwen3_5_moe_text"
• Weight keys at model.layers.* (standard causal LM format, no language_model. prefix)
• Vision encoder and MTP weights removed

Model structure

• Architecture: Hybrid GatedDeltaNet + Full Attention, Sparse Mixture-of-Experts
• Total parameters: ~35B (3B active per token)
• Dtype: bfloat16

How to use with vLLM

from vllm import LLM
llm = LLM(model="codecho/Qwen3.5-35B-A3B-text-only", trust_remote_code=True, tensor_parallel_size=2)