Parameter count, and what it fixes
A parameter count is the number of learned values in a set of weights. It fixes how large the download is and which variant is being discussed, and it is a poor predictor of the memory a run will consume. As of 2026-09-22.
| Factor | Effect |
|---|---|
| Numeric format | Two bytes per parameter in half precision, one in fp8 |
| Activations | Added at run time, and not in the count |
| Text encoder | Sometimes counted, sometimes not |
| Autoencoder | Sometimes published as a separate figure |
Inclusion rule. Factors that sit between a parameter count and a memory figure. A factor a vendor sometimes folds into the count is listed, because that is the ambiguity. Order. From the factor that acts on every parameter to the ones that are counted inconsistently.
1Counting conventions are not settled
Whether a text encoder or an autoencoder is inside the headline figure varies by publisher. One release in this register publishes its autoencoder separately, and another writes its total as a bound rather than a number.
That is why a count is useful mainly for placing a release in an order of magnitude, and why the register keeps the vendor's phrasing instead of normalising it.
2The arithmetic that looks convincing
Two bytes per parameter in half precision gives a weights footprint, and people stop there. Activations, the framework, the encoder and any offload strategy all sit outside that sum.
The register has a thirteen billion parameter release stating a six gigabyte floor and a two billion release requiring over thirty-two. Neither is explicable from the count alone.
3How this column is filled across the register
Six families publish a single round figure, four publish counts across several variants, two publish component breakdowns, one publishes an exact count, one publishes a bound and one publishes nothing.
Only the exact count and the round figures can be placed beside each other directly. A bound has no upper end, and a component split has to be read with its labels rather than summed.
A reading note, not an entry: no vendor value appears on this page. Where the register records this term for a particular release, it is on the size column and on that family's own page. Nearby terms: Checkpoint, Model variant.