Tip: Qwen 3.6 27B Fits in 16.8GB Q4_K_M — Flagship-Level Local Coding on a Single Consumer GPU

Simon Willison·medium signal

Simon Willison tested Qwen3.6-27B's Q4_K_M quantization (16.8GB) running locally via llama.cpp and confirmed it generates working code on consumer hardware including an RTX 4090 or M-series Mac. For builders wanting a local coding model without cloud API dependency, this is the first sub-30B model to match SWE-bench scores within 4 points of Claude Opus 4.6. Critical caveat: Qwen's benchmarks use their own internal agent scaffold — independent third-party reproduction outside that scaffolding remains limited as of April 23.