Alibaba Qwen Releases 27B Parameter Model That Claims to Match 397B Performance on Coding Tasks

Alibaba Qwen released Qwen3.6-27B, a 27B parameter dense model that claims to deliver "flagship-level agentic coding performance" surpassing their previous-generation Qwen3.5-397B-A17B (a 397B total parameter, 17B active MoE model) across all major coding benchmarks, according to the company.

The size difference is significant: Qwen3.6-27B is 55.6GB on Hugging Face compared to 807GB for Qwen3.5-397B-A17B. A quantized Q4_K_M version from Unsloth reduces the footprint to 16.8GB.

Performance Testing

Simon Willison tested the 16.8GB quantized version using llama.cpp's llama-server. In a test generating an SVG of a pelican riding a bicycle, the model produced a detailed, coherent image with correct bicycle geometry (spokes, chain, frame), a recognizable pelican, and background details including clouds, birds, and grass.

Performance metrics from llama-server:

• Prompt processing: 54.32 tokens/s (20 tokens in 0.4s)
• Generation speed: 25.57 tokens/s (4,444 tokens in 2min 53s)

Model Availability

Qwen3.6-27B is available as open weights on Hugging Face. The model includes reasoning mode support via the --reasoning on flag and uses a 65,536 token context window in testing configurations.

Specific benchmark scores and pricing were not disclosed in the announcement. The model represents Qwen's approach to achieving competitive coding performance in a significantly smaller architecture compared to their MoE models.

What This Means

If the coding benchmark claims hold up to independent verification, Qwen3.6-27B represents a substantial efficiency gain—achieving similar performance to a 397B parameter model in a 27B dense architecture. The 16.8GB quantized version running locally at 25 tokens/s makes flagship-level coding capabilities accessible on consumer hardware. However, the specific benchmarks and scores referenced in Qwen's "all major coding benchmarks" claim require independent validation.