Scores

What the numbers measure.

Two scores describe the repository. One estimate describes memory. None of them is a grade of the model's answers, and none of them is a run we timed. Records checked 2026-11-01.

Freshness is age

A repository opened today scores 100. The score halves every 180 days: 100 times 2 to the power of minus age in days, divided by 180. A repository opened 180 days ago scores 50. Missing dates stay unscored. The date is the day Hugging Face records for the repository, which can differ from a paper date.

Popularity is attention

Downloads carry 80 percent and likes carry 20. Both use a log scale, capped at 10 million downloads and 100,000 likes. The counts are whatever Hugging Face publishes on the model record. This desk does not recount them. A missing count is left out of its half of the score rather than treated as zero attention.

Fit is arithmetic

A 4-bit estimate uses 0.55 bytes per parameter. 8-bit uses 1.05. FP8 uses 1.0. FP16 uses 2.0. On top of that weight, the page adds a fixed allowance: 1 GB under 8 billion parameters, 1.5 GB under 20 billion, 2.5 GB under 40 billion, and 4 GB above that, for a text model. Vision adds 2 GB to the text allowance. An image model adds 4 GB. A video model adds 8 GB. Audio adds 0.5 GB under a billion parameters and 1 GB otherwise.

Worked examples, not measurements. A 7 billion parameter model at 4-bit plus the allowance is about 4.9 GB. On 24 GB that is fits. A 70 billion parameter model at 4-bit plus the allowance is about 42.5 GB. On 8 GB that is too-large.

The bands, compared in thousandths of a gigabyte so the threshold does not drift: fits at or under 72 percent of usable memory, tight at or under 95 percent, offload at or under 160 percent, and too large past that. Usable memory is the full figure for NVIDIA and for a rental, 70 percent on a Mac, and 45 percent for CPU-only.

Where the repository publishes separate weight files, the table also names a measured file. A GGUF repository contributes its smallest pack, not the sum of every pack. A checkpoint repository contributes its largest file, not a small encoder sitting beside it. Sharded weights are summed.

What we will not print

No benchmark trophy, no claim that a model runs because a formula says the weights could fit, and no rental price. If the public record has no parameter count and no usable file size, the model is left out rather than given a guessed size.

This directory is assembled from public model records. Freshness is age. Popularity is downloads and likes. A fit figure is an estimate from a parameter count or from a published file size, not a run on our hardware. How the numbers are made. AI disclosure.