Devoured - April 23, 2026
Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model (2 minute read)

Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model (2 minute read)

AI llmcoding Read original

Qwen's new 27B parameter model reportedly matches flagship-level coding performance in a package 14x smaller than their previous best model.

What: Qwen3.6-27B is a 27B dense model (55.6GB full size, 16.8GB quantized) that the Qwen team claims surpasses their previous 397B parameter mixture-of-experts flagship model on all major coding benchmarks while being significantly smaller and more accessible for local deployment.
Why it matters: This represents a significant efficiency leap where a compact model can deliver top-tier performance previously only possible with massive models, making powerful coding assistance practical to run locally on consumer hardware rather than requiring cloud infrastructure.
Takeaway: Download and run the quantized version locally using llama.cpp with the provided recipe to test flagship-level coding assistance on your own machine.
Deep dive
  • Qwen3.6-27B achieves what the team claims is flagship-level coding performance while being 807GB → 55.6GB in full size compared to their previous best model
  • The previous generation Qwen3.5-397B-A17B was a mixture-of-experts architecture with 397B total parameters and 17B active, while the new model is a 27B dense architecture
  • Simon Willison tested a 16.8GB quantized version (Q4_K_M) locally using llama-server and found it delivered impressive results on SVG generation tasks
  • The quantized model generated complex SVG images (pelican on bicycle) at 25.57 tokens/s for generation, producing 4,444 tokens in under 3 minutes on consumer hardware
  • Testing methodology used specific llama-server configuration with reasoning mode enabled and thinking preservation to maximize coding performance
  • The model downloaded to local cache (~17GB) on first run, making it practical for offline use once cached
  • Performance remained strong even with aggressive quantization, suggesting the model's capabilities are robust across different compression levels
  • SVG generation quality was described as "outstanding" for a 16.8GB local model, demonstrating practical coding assistance capabilities
  • The breakthrough suggests that architectural improvements and training advances are enabling smaller dense models to compete with much larger sparse models
Decoder
  • Dense model: A model where all parameters are used for every inference, unlike mixture-of-experts which activates only a subset
  • MoE (Mixture of Experts): Architecture with many parameter groups where only some are active per request, allowing larger total size with lower active memory
  • Quantized: Compressed model using lower precision numbers (like 4-bit instead of 16-bit) to reduce size while preserving most capability
  • GGUF: A file format for storing quantized models optimized for efficient inference on CPUs and GPUs
  • Agentic coding: AI systems that can autonomously plan, execute, and iterate on programming tasks rather than just generating code snippets
  • Q4_K_M: A specific quantization scheme using 4-bit precision with medium-size optimization for balancing quality and size
Original article

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

Big claims from Qwen about their latest open weight model:

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.

On Hugging Face Qwen3.5-397B-A17B is 807GB, this new Qwen3.6-27B is 55.6GB.

I tried it out with the 16.8GB Unsloth Qwen3.6-27B-GGUF:Q4_K_M quantized version and llama-server using this recipe by benob on Hacker News, after first installing llama-server using brew install llama.cpp:

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}'

On first run that saved the ~17GB model to ~/.cache/huggingface/hub/models--unsloth--Qwen3.6-27B-GGUF.

Here's the transcript for "Generate an SVG of a pelican riding a bicycle". This is an outstanding result for a 16.8GB local model:

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.

Performance numbers reported by llama-server:

  • Reading: 20 tokens, 0.4s, 54.32 tokens/s
  • Generation: 4,444 tokens, 2min 53s, 25.57 tokens/s

For good measure, here's Generate an SVG of a NORTH VIRGINIA OPOSSUM ON AN E-SCOOTER (run previously with GLM-5.1):

Digital illustration in a neon Tron-inspired style of a grey cat-like creature wearing cyan visor goggles riding a glowing cyan futuristic motorcycle through a dark cityscape at night, with its long tail trailing behind, silhouetted buildings with yellow-lit windows in the background, and a glowing magenta moon on the right.

That one took 6,575 tokens, 4min 25s, 24.74 t/s.