Pelican showdown

In the benchmark of a pelican riding a bicycle, Qwen3.6-35B-A3B outperformed Claude Opus 4.7. The author ran the 20.9GB Qwen3.6-35B-A3B-UD-Q4_K_S.gguf quantized model, provided by Unsloth, on a MacBook Pro M5 via LM Studio and the llm-lmstudio plugin.

On the first attempt, the pelican drawn by Qwen rendered even the bicycle frame stably, while Opus 4.7 broke the frame. Retrying Opus with thinking_level: max did not bring much improvement.

As a safeguard, the author also compared a "flamingo riding a unicycle" SVG, and judged Qwen to be better in this test as well. In particular, the author praised the SVG comment <!-- Sunglasses on flamingo! --> that appeared in Qwen's output.

However, the author emphasizes that this pelican benchmark was originally something close to a joke. In the past, this result tended to correlate fairly well with a model's overall usefulness, but this time that connection seems broken. It's hard to say a 21GB local quantized model is more capable than Anthropic's latest proprietary model, but the author concludes that, at least for now, if you want to make an SVG of a pelican riding a bicycle, Qwen is the better choice.