Home / Benchmark / swiss-style
Seven models, twenty-one runs, every one of them correct.
We include this style as a control, and it earns the place. Swiss Style is pure geometry and type: a strict grid, one typeface, hard edges. We gave seven models the same DESIGN.md and ran each three times. Every one of the twenty-one outputs matched the spec on shape, palette and type.
Every model, same spec, three runs
Shape fidelity is the share of border-radius and box-shadow values in the output that match the spec. Range is the spread between the best and worst of the three runs for that model. Cells below 75% are shaded. Seven open-weight models, each on its own vendor endpoint, no aggregator, temperature 0.7, thinking disabled, three independent samples.
| Model | Shape (median) | Range | Palette | Type |
|---|---|---|---|---|
| Qwen3.8-Max | 100.0% | 0pp | 100.0% | 100.0% |
| Doubao Seed 2.1 Turbo | 100.0% | 0pp | 100.0% | 100.0% |
| GLM-5.3 | 100.0% | 0pp | 100.0% | 100.0% |
| Qwen3.8-Flash | 100.0% | 0pp | 100.0% | 100.0% |
| Doubao Seed 2.1 Pro | 100.0% | 25pp | 100.0% | 100.0% |
| DeepSeek V4.1 Flash | 100.0% | 0pp | 100.0% | 100.0% |
| Kimi K3.1 Pro | 100.0% | 0pp | 100.0% | 100.0% |
What breaks
Three places where this style leaves the model something to decide. Each one is a gap in the specification, not a reading error by the model.
- border-radius — One value, used everywhere
The spec names a single radius and the style uses it on every surface. There is nothing to interpret. - box-shadow — Either none, or one declared shadow
No depth system to infer, so the shadow is a copy operation. - grid and spacing — A declared spacing scale
Every gap is a token. The model selects from a list; it never composes.
The finding
When the style is fully determined by discrete values, the models are reliable. Twenty-one runs, zero failures. The contrast with Illustration and Maximalism is the point: the difference is not model capability, it is how much the spec leaves open.
One caveat: Doubao Seed 2.1 Pro still spanned 25 points across its three runs, while holding a 100% median. A perfect median is not the same as a stable model — that is the reason we publish the range alongside the score.
Coverage and method. This page covers one of six benchmark styles, the only six for which all seven models were run. The other 22 styles in the library were measured on a single model (Doubao Seed 2.1 Turbo, the one run across all 28 styles), so their numbers describe that model’s variance and are not cross-model comparisons. Every figure here comes from a machine check of the generated HTML against the spec’s token table — no human rating, no screenshots. Raw requests and responses are archived; how it is measured.