Qwen3.6-35B released

Qwen3.6-35B-A3B has been released. It is the first open-weight version of Qwen3.6, a multimodal model with 35B total parameters and a 3B active MoE structure.

It handles text, images, and video together, with a default context length of 262,144 tokens that can be extended up to 1,010,000 tokens. It is distributed under the Apache 2.0 license and is compatible with Hugging Face Transformers, vLLM, SGLang, and KTransformers.

There are two key improvements.

• Agentic coding: Strengthened frontend workflows and repository-level reasoning
• Thinking preservation: Retains reasoning traces from past conversations to reduce repetitive work

Representative benchmark results are SWE-bench Verified 73.4, SWE-bench Pro 49.5, MMLU-Pro 85.2, AIME26 92.7, MMBench EN-DEV-v1.1 92.8, and VideoMMMU 83.7.