Qwen3.6-27B: 27B Model Claims to Beat 397B MoE on All Coding Benchmarks

Qwen3.6-27B: 27B Model Claims to Beat 397B MoE on All Coding Benchmarks

Alibaba's Qwen team released Qwen3.6-27B under Apache 2.0 — a dense 27B model claiming to outperform Qwen3.5-397B-A17B on every major coding benchmark, with community testers also claiming it beats Claude Opus 4.5. Unsloth Dynamic GGUFs bring the model down to a 16.8GB file runnable on 18GB RAM, and a community quantization (PRISM-NVFP4) delivers 120 tokens/second on RTX hardware. The model supports both thinking and non-thinking modes and passes SWE-Bench evaluations above MiniMax-M2.5.

Why It Matters

A 27B parameter model matching or beating much larger proprietary and open-weight competitors — while fitting on consumer hardware — accelerates the "end of subscription era" framing already circulating on X. If community benchmark claims hold, Qwen3.6-27B shifts the cost-performance frontier for local coding agents significantly. See the Qwen blog for full benchmark tables.