Tongyi Qianwen open-source Qwen3.6-27B, the 27B dense model's encoding ability surpasses the previous generation 397B flagship

ME News report: On April 22 (UTC+8), according to Beating Monitoring, the Alibaba Tongyi Qianwen team open-sourced Qwen3.6-27B, a 27B-parameter dense multimodal model focused on coding agent capabilities. This is the third member of the Qwen3.6 series, following the API version Qwen3.6-Plus and the small-activated MoE version Qwen3.6-35B-A3B, with its weights already released on Hugging Face and ModelScope.

The core selling point is that the 27B dense architecture fully outperforms the previous-generation open-source flagship Qwen3.5-397B-A17B (397B total parameters, 17B activated MoE model). In coding agent benchmarks, SWE-bench Verified scored 77.2 versus 76.2, SWE-bench Pro scored 53.5 versus 50.9, Terminal-Bench 2.0 scored 59.3 versus 52.5, and SkillsBench scored 48.2 versus 30.0. On reasoning tasks, GPQA Diamond scored 87.8, approaching the performance of models with several times the parameter count.

For visual agents, AndroidWorld scored 70.3, higher than Qwen3.5-27B’s 64.2. The model natively supports image and video inputs, and its thinking mode and non-thinking mode share the same set of weights. Since the dense architecture does not involve MoE routing, deployment is simpler than that of the 397B MoE. The official documentation shows it can directly connect to three terminal coding tools: OpenClaw, Claude Code, and Qwen Code. The API will be launched on Alibaba Cloud’s Bailing platform. (Source: BlockBeats)