27B Qwen3.6 Beats 397B MoE on Coding Benchmarks
Qwen3.6-27B dense model surpasses Qwen3.5-397B-A17B (397B total, 17B active MoE) on all major coding benchmarks while using 55.6GB vs 807GB; quantized 16.8GB version generates detailed SVGs locally at 25 tokens/s.
Compact Model Delivers Flagship Coding Power
Qwen3.6-27B achieves agentic coding performance matching proprietary flagships, outperforming the prior open-source leader Qwen3.5-397B-A17B across all major coding benchmarks. At 27B dense parameters, it totals 55.6GB on Hugging Face—over 14x smaller than the 807GB MoE predecessor (397B total, 17B active). This size efficiency enables local deployment without massive hardware, challenging the assumption that larger models always win on specialized tasks like coding.
Local Deployment Yields Production-Grade Outputs
Run the 16.8GB Q4_K_M quantized version from Unsloth via llama-server (installed via brew install llama.cpp) with this command for optimal reasoning and generation:
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}'
First run caches the model at ~/.cache/huggingface/hub/models--unsloth--Qwen3.6-27B-GGUF. Sampling params (temp 0.6, top-p 0.95, top-k 20) balance creativity and coherence for coding/generation tasks.
Real-World Generation Benchmarks
Prompt "Generate an SVG of a pelican riding a bicycle" produced 4,444 tokens in 2min 53s (25.57 tokens/s generation, 54.32 tokens/s reading)—yielding a detailed SVG with accurate bike mechanics (spokes, chain, frame), pelican anatomy (bill, wings on handlebars), and scenic background (clouds, birds, grass, sun). A follow-up "NORTH VIRGINIA OPOSSUM ON AN E-SCOOTER" generated 6,575 tokens in 4min 25s (24.74 tokens/s), creating neon Tron-style visuals with precise details like trailing tail and silhouetted cityscape. These outputs demonstrate production-ready creative coding from a local 16.8GB model, far exceeding typical small-model quality.