AI BriefingQwen3.6-27B: Flagship-Level Coding Realized with a 27B Dense Model
Key point
Qwen has open-sourced Qwen3.6-27B, a 27B dense multimodal model.
Details
Qwen has released Qwen3.6-27B as open weights. It is a 27B dense multimodal model that supports both thinking / non-thinking modes, positioning agentic coding performance as its core strength.
The main claim is better coding performance than the previous-generation open-source flagship, Qwen3.5-397B-A17B. The published benchmarks show the following figures.
- SWE-bench Verified: 77.2
- SWE-bench Pro: 53.5
- Terminal-Bench 2.0: 59.3
- SkillsBench Avg5: 48.2
- GPQA Diamond: 87.8
Vision-language performance was also disclosed. It supports multimodal reasoning across images and video, document understanding, visual question answering, spatial reasoning, video understanding, and visual agent tasks.
There are three deployment options.
- Immediate use in Qwen Studio
- Calling via the Alibaba Cloud Model Studio API
- Downloading weights from Hugging Face / ModelScope
It also supports the preserve_thinking feature, and integration examples with external coding agents such as OpenClaw, Claude Code, and Qwen Code are provided. It emphasizes that, being a dense architecture, it is easy to deploy without the routing complexity of MoE.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Privatemode Benchmarks GLM-5.3-Flash as Vision-Capable Decision Model with Parity to Jev but Higher Cost
Qwen Intelligence Launches Three Mobile AI Agents and Open Benchmark SuiteQwen3.5-based Small Decision Model 'Kev' Released with TypeSafe API Compatibility and Performance Metrics