Qwen Open-Sources Qwen3.6-35B-A3B Sparse MoE Model

Details

• Qwen released Qwen3.6-35B-A3B, a sparse Mixture-of-Experts (MoE) model with 35B total parameters and 3B active parameters, under Apache 2.0 license.
• Excels in agentic coding, matching performance of models 10x its active size, and outperforms dense Qwen3.5-27B on key coding benchmarks.
• Dramatically surpasses predecessor Qwen3.5-35B-A3B in agentic coding and reasoning tasks.
• Natively multimodal with strong vision-language perception and reasoning, exceeding expectations for its size across most benchmarks.
• Features multimodal thinking and non-thinking modes, supporting unified vision-language foundation for reasoning, coding, agents, and visual understanding.
• Builds on Qwen3.5 architecture with gated delta networks and sparse MoE, enabling high-throughput inference and 256K native context length.

Impact

Qwen3.6-35B-A3B advances open-source AI by delivering frontier-level agentic coding and multimodal reasoning with just 3B active parameters, outpacing its 22B-active predecessor Qwen3-235B-A22B and rivaling dense models like Qwen3.5-27B that match GPT-5 mini on SWE-bench. This efficiency pressures closed rivals like OpenAI and Anthropic by lowering inference costs and enabling deployment on consumer hardware, widening access to high-performance vision-language agents. It narrows the gap in video understanding and benchmarks like VideoMME where larger models like Gemini 3 Pro lead, accelerating adoption in resource-constrained environments.