Alibaba Releases Qwen3.6-27B - Open Source Coding That Matches Claude 4.5 Opus

Alibaba's Qwen team has released Qwen3.6-27B, a dense 27-billion parameter open-source coding model that challenges conventional wisdom about model scaling. The model outperforms Alibaba's own 397-billion parameter Mixture-of-Experts model across every major coding benchmark, despite being nearly 15 times smaller. On SWE-bench Verified, Qwen3.6-27B achieves a score of 77.2, surpassing its larger predecessor. Perhaps more impressively, it matches Claude 4.5 Opus exactly at 59.3 on Terminal-Bench 2.0, demonstrating competitive performance against one of the industry's leading proprietary models. The model features a hybrid architecture that combines Gated DeltaNet linear attention with traditional self-attention mechanisms. A new 'Thinking Preservation' feature maintains reasoning context across tool calls, enhancing the model's agentic coding capabilities. This architectural innovation appears to compensate for the smaller parameter count whilst delivering superior performance. Released under the Apache 2.0 licence, Qwen3.6-27B can be freely used and modified by developers and organisations. The open-source nature, combined with its relatively modest size, makes it accessible for deployment on local hardware, potentially enabling widespread adoption across the developer community.