Qwen3.6-27B: Flagship-Level Coding Performance from a 27B Dense Model

Qwen3.6-27B was released as a 27-billion-parameter dense multimodal model, supporting both thinking and non-thinking modes from a single unified checkpoint. It also handles image and video processing, targeting both text and multimodal reasoning.

The core highlight is agentic coding performance. On major coding benchmarks, it surpassed the previous-generation open-source flagship Qwen3.5-397B-A17B, and is claimed to outperform models with up to 15x more total parameters.

Key figures are as follows.

• 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
• SkillsBench 48.2 vs 30.0
• GPQA Diamond 87.8, AIME26 94.1

Beyond coding, knowledge, STEM, and reasoning performance were also disclosed. For example, it recorded MMLU-Pro 86.2, MMLU-Redux 93.5, SuperGPQA 66.0, and C-Eval 91.4, while STEM and reasoning items showed results such as LiveCodeBench v6 83.9, HMMT Feb 25 93.8, and IMOAnswerBench 80.8.

On the vision-language side, the model was evaluated on document understanding, spatial intelligence, video understanding, and visual agent tasks. Examples include results such as MMMU 82.9, OCRBench 89.4, VideoMME(w sub.) 87.7, and AndroidWorld 70.3.

In terms of architecture, dense was chosen to simplify deployment without MoE routing complexity. Qwen emphasizes that this model delivers top-tier coding performance at a practically deployable scale.

Access paths are also broad.

• Open weights: Hugging Face, ModelScope
• API: Alibaba Cloud Model Studio
• Qwen Studio: available for immediate trial
• Integrations: OpenClaw, Qwen Code, Claude Code support

On the API side, this release supports the preserve_thinking feature, described as being used to preserve thinking content from previous turns in agentic tasks. Model Studio also provides an OpenAI-compatible chat completions/responses API and an Anthropic-compatible interface.