SVG Model Match

A blind preference and model-identification game built from 40 locally generated SVG illustrations: eight unusual prompts rendered by five open-weight models on an RTX 4090.

Models:
- Qwen 3.6 27B, thinking enabled through Ollama
- Qwen 3.8 27B, Unsloth UD-Q5_K_XL GGUF, xhigh reasoning through Ollama
- Muse Glimmer 30B official K-Quant 17GB, xhigh reasoning, DFlash
- Gemma 4 26B A4B, thinking enabled
- Gemma 4 31B dense, thinking enabled

The game first records a blind favourite, then asks the player to match each image to its model. Model identification is scored over 40 assignments. Source SVGs are served alongside the artifact.

This is a playful perceptual test, not a controlled benchmark. Sampling used each model's recommended settings; some source runs required retries to obtain complete SVGs, and the 31B Gemma run intentionally removed an earlier reasoning-length instruction. Qwen 3.8 used xhigh reasoning with temperature 1.0, top_p 0.95, top_k 20, min_p 0, presence penalty 0 and repetition penalty 1.0. Seven scenes used a 40,960-token context/output budget; the crocodile used 65,536 after twice exhausting the smaller budget. Each displayed Qwen 3.8 scene description was prefixed with “Generate an SVG of”; its SVG was extracted without editing and XML-validated before publication.