Available model measurements

What the figures can tell us

Cumulative popularity does not measure recent growth. A benchmark depends on its tasks, date and model coverage. Scores are relative to the observed pool.

The team's opinions are signed and separate from the calculation. A Talk Show recommendation does not change a rank.

This archive keeps its historical version. The v3 method and its calibration are tracked separately and are not applied retroactively.

Preserved measurements and sources

The measurements below belong to this edition. A metric shown for context does not necessarily contribute to the score. Effective weights are shown only where they were preserved.

Historical configuration

The parameters below were frozen at calculation time. They may contain planned dimensions absent from some rows. They do not replace the preserved measurements and effective weights.

{
  "size": 20,
  "valueWeights": {
    "price": 0.4,
    "quality": 0.6
  },
  "qualityWeights": {
    "arenaText": 0.5,
    "artificialAnalysis": 0.5
  },
  "minimumQualityCandidates": 20
}

Open this edition's JSON data ↗ (new tab)

Observation, calculation and publication

An observation dates a measurement at a source. Calculation freezes the score using available observations. Publication makes the edition accessible. These dates can differ; a new collection does not automatically republish a ranking.

Scheduled calculation slots at 03:17 and 15:17 UTC do not guarantee a new edition: coverage, baselines and validation can delay it. The displayed date belongs to the edition actually published.