Build-off · SVGBench · Design skills

Comparing tasks between Qwen3.8 Flash Next and frontier models

This was made for fun, without much scientific rigour — the artifacts will show you that. Do not take it too seriously, but I think it shows a good model running locally, and at a very good speed.

Its game and its drawings sit next to what hosted frontier models made from the same prompts, and they hold up. Judge them by eye: the scores are in each section, and they say less than the artifacts do. Every picture is the model's own work — the SVG file it wrote, or a screenshot of the app it built. A stack is one harness running one model.

Versions of a web runner game

Ability to draw SVGs

Each row heading is the verbatim prompt; each drawing is one stack's answer. Scored drawings show requirements met over requirements asked, judged by claude-opus-5 on the rendered drawing. Drawings without a verdict are marked unscored and excluded from the mean. Open a scored drawing to read every requirement with the judge's note. Last take is the drawing a stack ended on, preferring an animated one; best take is its highest-scoring drawing on that prompt. Unscored stacks do not appear in best take. Take evolution lines up one stack's takes of one drawing in the order it made them, with the change in requirements met at each step.

Show

Ability with layout

Group