HeadsUpAI

Local Qwen3.6 Outperforms Claude Opus 4.7 on Complex Spatial Coding Tasks

· Updated

Simon Willison, creator of Datasette, reported that Alibaba's new Qwen3.6-35B-A3B open-weight model outperformed Anthropic's flagship Claude Opus 4.7 in SVG generation tests. A 21GB quantized version produced superior spatial results for prompts that the frontier model failed to render correctly.

This result challenges the assumption that frontier models always provide the highest quality for technical tasks. Even when Claude Opus 4.7 used its maximum thinking budget, it failed to correctly render a bicycle frame that the smaller Qwen model handled accurately. Scale and thinking tokens do not always guarantee superior spatial reasoning.

Run Qwen3.6-35B-A3B locally using tools like LM Studio or Ollama with GGUF (a file format optimized for local LLM execution) weights. This setup is effective for SVG illustration and technical code generation. The model is available on Hugging Face as a cost-effective alternative to proprietary APIs.

Simon Willison
Simon Willison
@simonw
X

Shocking result on my pelican benchmark this morning, I got a better pelican from a 21GB local Qwen3.6-35B-A3B running on my laptop than I did from the new Opus 4.7! Qwen on the left, Opus on the right https://t.co/kDlbnJv6YI

130retweets2.1klikes
View on X

Share this update