Qwen3.6-27B Dense Model Surpasses Flagship MoE in Coding Benchmarks

Qwen has launched a new model called Qwen3.6-27B, which boasts 27 billion parameters and outperforms the previous open-source champion, Qwen3.5-397B-A17B, with its 397 billion parameters and 17 billion active MoE, in crucial coding tests. This latest model is much more efficient, taking up only 55.6GB on Hugging Face compared to the 807GB of its predecessor. Simon Willison tested the 16.8GB quantized version, Qwen3.6-27B-GGUF Q4_K_M, using llama-server and a technique from benob on Hacker News. It managed to create an SVG of a pelican on a bike, while another SVG of a North Virginia opossum on an e-scooter took 6,575 tokens and 4 minutes 25 seconds to generate, achieving 24.74 tokens per second. This was announced on April 22, 2026.