Qwen3.6-27B tops giant predecessor on coding

This is Qwen’s clearest pitch yet for dense local coding models: fewer architectural tricks, more predictable deployment, and benchmark numbers close enough to make teams question whether they need huge MoE models for everyday agent work.

• –The headline claim is strong: 77.2 on SWE-bench Verified, 53.5 on SWE-bench Pro, and 59.3 on Terminal-Bench 2.0, beating Qwen3.5-397B-A17B on those coding evals.
• –Dense 27B matters because quantized deployments can be far simpler than serving sparse MoE checkpoints, especially for local coding agents and self-hosted developer workflows.
• –Multimodal support, 128K context in the OpenClaw config, and thinking-preservation support make this more than a code-completion model; Qwen is aiming at full repo agents.
• –Integration notes for OpenClaw, Qwen Code, and Claude Code-compatible APIs show the real battleground is agent runtime compatibility, not just raw leaderboard placement.