Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model

<p><strong><a href="https://qwen.ai/blog?id=qwen3.6-27b">Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model</a></strong></p> Big claims from Qwen about their latest open weight model:</p> <blockquote> <p>Qwen3.6-27B delivers flagship-level agentic coding performance, surpassing the previous-generation open-source flagship Qwen3.5-397B-A17B (397B total / 17B active MoE) across all major coding benchmarks.</p> </blockquote> <p>On Hugging Face <a href="https://huggingface.co/Qwen/Qwen3.5-397B-A17B/tree/main">Qwen3.5-397B-A17B</a> is 807GB, this new <a href="https://huggingface.co/Qwen/Qwen3.6-27B/tree/main">Qwen3.6-27B</a> is 55.6GB.</p> <p>I tried it out with the 16.8GB Unsloth <a href="https://huggingface.co/unsloth/Qwen3.6-27B-GGUF">Qwen3.6-27B-GGUF:Q4_K_M</a> quantized version and <code>llama-server</code> using this recipe by <a href="https://news.ycombinator.com/item?id=47863217#47865140">benob on Hacker News</a>, after first installing <code>llama-server</code> using <code>brew install llama.cpp</code>:</p> <pre><code>llama-server \ -hf unsloth/Qwen3.6-27B-GGUF:Q4_K_M \ --no-mmproj \ --fit on \ -np 1 \ -c 65536 \ --cache-ram 4096 -ctxcp 2 \ --jinja \ --temp 0.6 \ --top-p 0.95 \ --top-k 20 \ --min-p 0.0 \ --presence-penalty 0.0 \ --repeat-penalty 1.0 \ --reasoning on \ --chat-template-kwargs '{"preserve_thinking": true}' </code></pre> <p>Here's <a href="https://gist.github.com/simonw/4d99d730c840df594096366db1d27281">the transcript</a> for "Generate an SVG of a pelican riding a bicycle". This is an <em>outstanding</em> result for a 16.8GB local model:</p> <p><img alt="Bicycle has spokes, a chain and a correctly shaped frame. Handlebars are a bit detached. Pelican has wing on the handlebars, weirdly bent legs that touch the pedals and a good bill. Background details are pleasant - semi-transparent clouds, birds, grass, sun." src="https://static.simonwillison.net/static/2026/Qwen3.6-27B-GGUF-Q