Local Qwen3.6-35B Beats Claude Opus on SVG Pelicans
Quantized 20.9GB Qwen3.6-35B-A3B on an M5 MacBook Pro generates anatomically superior SVG pelicans riding bicycles—and charismatic flamingos on unicycles—compared to Anthropic's Claude Opus 4.7.
Run Quantized Open Models Locally for Creative SVG Tasks
To evaluate SVG generation capabilities, prompt models with 'Generate an SVG of a pelican riding a bicycle.' A 20.9GB Q4_K_S.gguf quantization of Alibaba's Qwen3.6-35B-A3B—loaded via LM Studio and the llm-lmstudio plugin on a MacBook Pro M5—produces a coherent pelican with proper bicycle proportions, outperforming Claude Opus 4.7, which distorts the bike frame. Even with Opus's thinking_level: max parameter, results remain flawed. This setup enables local inference of a 35B-parameter model without cloud dependency, ideal for iterative creative testing.
For validation against benchmark overfitting suspicions, test 'Generate an SVG of a flamingo riding a unicycle.' Qwen delivers a stylish flamingo with sunglasses, bowtie, cigarette, heart emojis, and flair-rich details (despite slightly long spokes), plus an excellent SVG comment. Opus yields a competent but dull, flairless illustration with black wheels. Qwen wins both rounds, proving robustness on unpublished prompts.
Niche Benchmarks No Longer Track Overall Utility
Historically, pelican SVG quality loosely correlated with model prowess: early 2024 outputs were junk, while recent ones (e.g., Gemini 3.1 Pro) became production-usable. This breaks the pattern—a local 21GB quantized open model now surpasses Anthropic's flagship proprietary release on this task, despite Opus's superior general utility. Use such joke benchmarks to spotlight specific strengths like SVG anatomy and charisma, but don't extrapolate to broader capabilities. For pelican SVGs specifically, prioritize local Qwen over Opus.