Pattern: Local Qwen3.6-35B Beats Opus 4.7 on Visual Tasks — Frontier Isn't Always Best

Simon Willison·medium signal

Simon Willison demonstrated that Alibaba's Qwen3.6-35B-A3B running locally on a laptop produced a better pelican drawing than Claude Opus 4.7 — on the same day Opus 4.7 launched with its 98.5% visual acuity benchmark. This is a recurring pattern: frontier models dominate aggregated benchmarks but specific open-weight models can outperform on narrow tasks at zero marginal cost. For builders: always test your actual use case against local models before committing to API spend.