Qwen3.6-35B-A3B (MoE 架构, 35B 总参数, 3B 激活参数)

Qwen3.6-35B-A3B represents a landmark release in efficient open-source AI. As the first open-weight model from the Qwen3.6 family, it employs a Mixture-of-Experts architecture with 35 billion total parameters but only activates approximately 3 billion per token—delivering inference costs and speeds comparable to a 3B dense model while achieving benchmark parity with 27B+ dense architectures. Released April 16, 2026 under Apache 2.0, it is available through Hugging Face, ModelScope, Alibaba Cloud Model Studio (as qwen3.6-flash), and the Qwen Studio playground.

The model's defining strength is agentic coding. On Terminal-Bench 2.0, it scores 51.5—a 27% improvement over its predecessor Qwen3.5-35B-A3B (40.5) and surpassing the much larger Gemma4-31B (42.9). SWE-bench Verified reaches 73.4%, nearly matching the 27B dense model's 75.0% while operating at a fraction of the compute. The QwenWebBench Elo rating surged from 978 to 1397, reflecting substantial gains in front-end code generation. Beyond coding, the model demonstrates remarkable multimodal capability: RealWorldQA (85.3) and OmniDocBench (89.9) both exceed Claude Sonnet 4.5, and spatial intelligence metrics like RefCOCO (92.0) and ODInW13 (50.8) set new benchmarks for models of this scale.

Architecturally, Qwen3.6-35B-A3B uses a 256-expert sparse MoE design where a router selects 9 experts per token (8 routed + 1 shared). This decouples model capacity from compute cost, allowing the 35B parameter storage to encode far more knowledge than a 3B dense model could hold. The model supports both 'thinking' (chain-of-thought reasoning) and 'non-thinking' (direct response) modes, with a critical 'preserve_thinking' feature that maintains reasoning context across multi-turn agentic workflows. Deployment is supported via vLLM, SGLang, and Transformers, with FP8 quantization reducing VRAM requirements to approximately 10GB.