Alibaba released Qwen3.6-27B. And this little model destroys their previous open-source monster Qwen3.5-397B-A17B across all agentic programming benchmarks. A model that is 15 times smaller overall scores 77.2 on SWE-bench Verified (vs 76.2 for 397B) and 59.3 on Terminal-Bench 2.0.

Features:
▫️ Under the hood, it retains native multimodality (understands images and videos)
▫️ Supports both thinking and non-thinking modes
▫️ Their API now supports the Anthropic protocol.

The last point means you can now set ANTHROPIC_BASE_URL to their endpoint and use Qwen3.6-27B under the hood of the Claude Code utility in the terminal. Or plug it into OpenClaw / Qwen Code and get a full-fledged coding agent.

Weights are here: Hugging Face