Qwen 3.6 27B: Dense 27B Model Beats 397B MoE on Agentic Coding Benchmarks

HuggingFace·high signal

Alibaba released Qwen3.6-27B on April 22, a dense open-weight model scoring 77.2% on SWE-bench Verified (within 3.7 points of Claude Opus 4.6) and 59.3% on Terminal-Bench 2.0, matching Claude 4.5 Opus exactly. On SkillsBench it hits 48.2% versus the previous 397B MoE's 30.0% — a 77% improvement with 14.8x fewer parameters. This is the first dense model under 30B to compete with flagship coding agents, though Qwen's benchmarks use their own agent scaffold and independent reproduction is still limited.