Four of seven models could not hold this style. One model failed 45 points apart between runs.
Maximalism stacks elements: many radii, overlapping surfaces, accents competing for the same corner. We gave seven models the same DESIGN.md and ran each three times. Four landed below 75% shape fidelity. The interesting part is not which models failed — it is that the failures are not reproducible between runs of the same model.
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 | 44.4% | 6pp | 100.0% | 100.0% |
| GLM-5.3 | 100.0% | 0pp | 100.0% | 100.0% |
| Qwen3.8-Flash | 54.5% | 12pp | 100.0% | 100.0% |
| Doubao Seed 2.1 Pro | 50.0% | 19pp | 100.0% | 100.0% |
| DeepSeek V4.1 Flash | 25.0% | 30pp | 85.7% | 100.0% |
| Kimi K3.1 Pro | 84.6% | 45pp | 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 — How many radii, and on what
Maximalism needs a radius system — base, raised, inset, pill. A spec that names one base radius leaves the model to invent the other three, and it will invent them differently each run. - box-shadow — Stack depth
The style depends on layered depth. Nothing in a token list tells a model how many layers to draw, so the count becomes a free parameter. - z-index and overlap — Which surface wins
When everything overlaps, the stacking order has to be specified. The spec does not, so it varies by run. This is the largest single contributor to the range we measured.
The finding
Maximalism is unstable rather than unreadable. The same model, the same spec, three runs: Kimi K3.1 Pro spanned 45 points. That is not a model that cannot do the style; it is a model whose output on this style is a coin flip.
This is the one style where we would not trust a single run — of any model, including the ones that scored 100%.
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.