Qwen open-sources Qwen3.6-35B-A3B coding model
Alibaba's Qwen team has released open weights for Qwen3.6-35B-A3B, a mixture-of-experts model that uses 3 billion active parameters and comes close to larger dense models on agentic coding benchmarks.
The Qwen team announced the release of Qwen3.6-35B-A3B as open weights. The sparse mixture-of-experts model has 35 billion total parameters, of which 3 billion are active. It follows the launch of Qwen3.6-Plus. The model is also available on Qwen Studio and through Qwen's API. It is natively multimodal and supports both thinking and non-thinking modes.
On the team's own evaluations, the model beats its predecessor, Qwen3.5-35B-A3B, on most agentic coding tasks. It scores 73.4 on SWE-bench Verified, up from 70.0, and 51.5 on Terminal-Bench 2.0, up from 40.5. On SWE-bench Pro it reaches 49.5, compared with 44.6 for the predecessor. Against the dense Qwen3.5-27B, results are mixed. The new model leads on Terminal-Bench 2.0, NL2Repo (29.4 vs. 27.3) and MCPMark (37.0 vs. 36.3), but trails on SWE-bench Verified (73.4 vs. 75.0) and several general-agent tests. Google's Gemma4-31B scored lower on most of the coding rows.
Qwen says the model's vision-language results match Claude Sonnet 4.5 on most benchmarks. It cites scores of 92.0 on RefCOCO and 50.8 on ODInW13 for spatial tasks.
All results are self-reported. Some come from internal benchmarks, including QwenClawBench and QwenWebBench. Qwen also used its own agent scaffold for the SWE-bench tests and ran SWE-bench Pro on a version of the public set with some tasks corrected. The announcement does not give licence terms or details on how the model was trained.