Qwen3.6-27B: Dense 27B Model Outperforms Own 397B MoE on SWE-Bench, Fits Single Consumer GPU

Qwen AI·medium signal

Alibaba's Qwen team released Qwen3.6-27B on April 22 under Apache 2.0 — a fully dense 27B multimodal model that scores 77.2% on SWE-bench Verified versus the previous flagship Qwen3.5-397B-A17B's 76.2%. The 397B model is 807GB on Hugging Face; the 27B replacement is 55.6GB and runs at Q4_K_M quantization (16.8GB) on a single consumer GPU. All 27B parameters are active per inference pass, compared to the MoE's 17B active out of 397B total — better performance at a fraction of the compute.