Qwen3.6-27B: Flagship-Level Coding Performance Achieved with a 27B Dense Model

Qwen claims that Qwen3.6-27B delivers flagship-level performance in agentic coding, stating it outperforms the previous generation open-source flagship Qwen3.5-397B-A17B across major coding benchmarks.

The difference in model size is also significant. According to Hugging Face, Qwen3.5-397B-A17B is 807GB, while the new model Qwen3.6-27B is 55.6GB.

In a hands-on test, the 16.8GB quantized version of Unsloth Qwen3.6-27B-GGUF:Q4_K_M was run with llama-server, installed via brew install llama.cpp, using the following settings:

• -c 65536
• --cache-ram 4096 -ctxcp 2
• --jinja
• --temp 0.6
• --top-p 0.95
• --top-k 20
• --reasoning on
• --chat-template-kwargs '{"preserve_thinking": true}'

After the first run, the model was stored at ~/.cache/huggingface/hub/models--unsloth--Qwen3.6-27B-GGUF. In terms of performance, it produced fairly impressive results on the "Generate an SVG of a pelican riding a bicycle" task, and the generation quality was rated excellent for a 16.8GB local model.

The numbers reported by llama-server were also solid. The read stage was 20 tokens / 0.4s / 54.32 tokens/s, and the generation stage was 4,444 tokens / 2min 53s / 25.57 tokens/s. Additionally, for the "NORTH VIRGINIA OPOSSUM ON AN E-SCOOTER" prompt, it recorded 6,575 tokens, 4min 25s, and 24.74 t/s.