Model Cards

Benchmark results come from first-party system cards, model cards, launch pages and evaluation reports. Each result is stored with its model, benchmark name and version, score, source location and reported test conditions.

Dedicated system and model cards form the canonical disclosure corpus. Launch pages and evaluation reports can support a capability result, but they do not count towards system-card length, evaluation totals or canonical-card overlap.

A result appears only when the score can be located in an official source, assigned to the correct model and tied to the benchmark label and qualifiers reported beside it. Qualifiers such as tool access, effort level, subset and harness are retained because they can materially change a result.

These are published results, including some attributed to external evaluators. Free Systems has not rerun the evaluations.

Comparisons follow a fixed order. We first look for the same benchmark with a sufficiently aligned setup. We then look for the same benchmark under different conditions, a related benchmark from the same family, and finally different benchmarks covering the same capability category.

The interface labels those relationships as direct comparison, qualified comparison, same benchmark with a different setup, related benchmark or different evidence. Only the first two support a score comparison. The others show how the labs chose to measure a capability; they do not establish which model is better.

Disclosure length is the word count of the canonical system or model card. PDF counts use extracted document text; HTML counts use the saved article or model-card body and exclude navigation, scripts and page furniture. Each source is tied to a stored document hash so the count refers to a specific version.

A system card covering multiple models retains its full document count for each model. We do not divide a shared document into estimated model-level word counts.

CAP is our subjective overall rating—not a benchmark score. We assign it editorially; it is not a standardized measurement or a calculated average of the bars. The creature artwork is illustrative.

The source material is uneven and changes over time. Labs use different benchmarks, versions, scaffolds and reporting conventions; many results cannot be reproduced from the published information alone. This is a working edition, and corrections will be recorded as the corpus is updated.