Model data and methodology

A reproducible finding

Of 6 selected language checkpoints, 2 fall at or below 32 GB and 3 at or below 64 GB under the same generic Q4 planning formula. 1 has no defensible simplified estimate. These are counts of this curated sample, not successful local runs or a census of all open models.

Reproduce by filtering the JSON on task = language and reasoning, then counting non-null memoryEstimateGB at each threshold. The rows link directly to publisher cards and individual profiles.

Fields and units

id

Stable catalog slug, not the publisher repository identifier.

parametersB / parameterBasis

Billions of stated model parameters or repository tensor elements, as qualified by the basis field; null when not verified. Auxiliary components may be separate.

score

LiveBench overall 0–100 for language, or modality-specific arena Elo for media; scoreLabel and scoreSource identify the scale. Never compare across scales.

memoryEstimateGB

Generic Q4 planning estimate in decimal GB for selected language checkpoints; not measured VRAM. Null means unknown or an inapplicable simplified formula.

repositoryStorageGB

Full published repository size where verified. It may include multiple formats and is not the minimum download.

Reviewed checkpoints