Qwen open-sources Qwen3.6-27B dense coding model
Alibaba's Qwen team has released open weights for Qwen3.6-27B, a dense multimodal model it says beats its own 397B-parameter previous-generation flagship on every major coding benchmark, at a size that is much easier to deploy.
The Qwen team announced Qwen3.6-27B. It is a dense 27-billion-parameter model released as open weights and is also available on Qwen Studio and through the Qwen API. It follows the earlier Qwen3.6-Plus and Qwen3.6-35B-A3B releases. A single checkpoint handles text, images and video, with both thinking and non-thinking modes.
The headline claim is coding performance. By Qwen's own numbers, the model beats Qwen3.5-397B-A17B, a mixture-of-experts model with 397B total and 17B active parameters, on SWE-bench Verified (77.2 vs. 76.2), SWE-bench Pro (53.5 vs. 50.9), Terminal-Bench 2.0 (59.3 vs. 52.5) and SkillsBench (48.2 vs. 30.0). It also scores 87.8 on GPQA Diamond.
Qwen's table includes Claude 4.5 Opus, which stays ahead on SWE-bench Verified (80.9) and SWE-bench Pro (57.1). Qwen3.6-27B ties it on Terminal-Bench 2.0 at 59.3 and edges past it on SkillsBench. Against Google's Gemma4-31B, the closest dense model in the comparison, Qwen reports large leads on the agentic coding tests. On knowledge benchmarks and HLE, the 27B model falls short of the older 397B flagship.
These are vendor-reported results using Qwen's internal agent scaffolds. Qwen says it corrected "some problematic tasks" in the public SWE-bench Pro set and re-ran all baselines on that revised version. SkillsBench was run on a 78-task subset. QwenWebBench and QwenClawBench are internal benchmarks.
Qwen's pitch is that a dense model avoids the routing complexity of mixture-of-experts models when deployed. The announcement does not include licensing terms or API pricing.