Local Qwen Model Outperforms Claude Opus 4.7 on Visual Benchmark
Simon Willison found a 21GB Qwen model running on a laptop beat Opus 4.7 at generating a pelican SVG.
Simon Willison (simonw) reported Thursday that Qwen3.6-35B-A3B, a 21GB model running locally via llama.cpp on his laptop, produced a better result than Claude Opus 4.7 on his "pelican benchmark" — a creative SVG generation test. He later confirmed the result held on a second test (a flamingo riding a unicycle), with full details and image comparisons posted to his blog.
The result is notable because Qwen3.6-35B-A3B is a mixture-of-experts model that activates only 3.5B parameters per forward pass, making it feasible to run on machines with as little as 32GB of RAM. That a locally-hosted open-weights model can match or exceed a frontier commercial model on visual reasoning tasks is a data point practitioners will watch: it bears directly on build-vs-buy decisions, data privacy trade-offs, and the ongoing question of how quickly open-weight capability is closing the gap with proprietary APIs.