Qwen releases 27B dense model claiming coding performance comparable to 397B predecessor

Qwen announced Qwen3.6-27B, a 27-billion parameter dense model designed for coding tasks. The company claims the model delivers coding performance matching its predecessor Qwen3.5-397B-A17B—a 397-billion parameter mixture-of-experts model—across major coding benchmarks. The new model is available on Hugging Face at 55.6GB in full precision, a significant size reduction from the 807GB predecessor.

The model has been quantized for local inference; Willison tested a 16.8GB Q4_K_M quantized version using llama.cpp's server implementation. In testing, the quantized model achieved reading speed of 54.32 tokens per second and generation speed of 25.57 tokens per second on an SVG generation task involving 4,444 output tokens. A second task generating 6,575 tokens completed at 24.74 tokens per second.

Willison's testing focused on qualitative output quality rather than standardized benchmarks. The source does not provide the specific coding benchmarks used to substantiate Qwen's performance claims, nor independent verification of those results. The comparison between dense and MoE architectures of significantly different sizes introduces architectural variables beyond simple parameter count.