Which models can you actually run?

Reach, by runtime

declared on the model cardnot declared, but ships a format it reads

Every bar but one is almost entirely light, which is the finding: for most runtimes the evidence is the file, not the card. MLX is the exception — 11,788 cards name it against 471 models shipping an npz, because mlx-community publishes converted weights under a name that says MLX rather than in the format that proves it. It is the one runtime where the card is the better evidence.

The most-used models candle can load

Either signal

182,092

Declared

0 — nobody writes it on a card

By format

182,092

Commercial use ok

111,698

Quantised build

95,655

Ordered by downloads, which on these 202,261 models is a real signal — 96% have a non-zero count. Parameter counts are resolved, not just read: the card first, then the base model's card, then the number in the name, then the GGUF file size, which is a per-parameter figure. That sizes 178,356 of the 182,092 models candle can load; the source is printed under every count that did not come from the card itself. Quantisation is read the same way: the card's quant_bits or method where stated, otherwise the GGUF builds the repository actually ships, which is where a range like 2–16-bit and a build count come from.