Qwen3.6-27B Outperforms Previous 397B Model in Coding Agent Benchmarks

ME News reports that on April 22 (UTC+8), according to monitoring by Beating, the Alibaba Tongyi Qianwen team has open-sourced Qwen3.6-27B, a 27-billion-parameter dense multimodal model optimized for coding agents. This is the third member of the Qwen3.6 series, following the API version Qwen3.6-Plus and the sparse MoE version Qwen3.6-35B-A3B, with weights now available on Hugging Face and ModelScope. Its key selling point is surpassing the previous open-source flagship, Qwen3.5-397B-A17B (397B total parameters, 17B activated MoE model), across all benchmarks using a much smaller 27B dense architecture. On coding agent benchmarks: 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. On reasoning tasks, it achieves 87.8 on GPQA Diamond—approaching the performance of models with several times more parameters. For visual agents, it scores 70.3 on AndroidWorld, outperforming Qwen3.5-27B’s 64.2. The model natively supports image and video inputs, with shared weights between thinking and non-thinking modes. The dense architecture eliminates MoE routing, simplifying deployment compared to the 397B MoE model. Official documentation confirms direct integration with three terminal coding tools: OpenClaw, Claude Code, and Qwen Code. An API version will be available on Alibaba Cloud’s Bailian platform. (Source: BlockBeats)