Which vendors publish an exact parameter count
One of the 15 families publishes a parameter count to the unit. One publishes a bound with no upper end, two publish component breakdowns, four publish counts across several variants, six publish a single round figure and one publishes nothing. As of 2026-09-12.
| Item | What is published | Detail |
|---|---|---|
| An exact count | One family | 2,059,174,912 parameters, published on the model card |
| A bound | One family | Over 13 billion parameters, with no upper end stated |
| Components published apart | Two families | VAE 175M and DiT 2.8B; 10 billion plus a 362M video autoencoder |
| Counts across variants | Four families | 2B and 5B; 1.3B, 5B and 14B; 24B and 4.5B; 13B with distilled and fp8 builds |
| A single round figure | Six families | 30 billion, 11B, 13B, 2B, 13.6B and a 5B |
| No count at all | One family | The card states clip lengths and resolutions instead |
Inclusion rule. One row per shape the count takes. A shape used once keeps its row, because a single exact count is what makes the rest of the column legible as approximations. Order. From the most precisely stated shape to the absence of one.
1Exactness removes the convention question
A rounded two billion could include or exclude a text encoder and a reader would never know. A count to the unit is a statement about one artefact, which is why one row here serves as the reference the rest of the column is read against.
It does not make the row more useful in every way. The same card still describes one configuration, and precision does not extend a figure to a different clip length or frame size.
2A bound is defensible and uncomparable
Counting conventions differ between publishers, and writing a bound avoids committing to one. That is reasonable for an author and it removes the row from the one comparison the size column exists to support.
The register keeps the phrasing. Rounding a bound into a number would manufacture a figure, and treating it as an exact count would manufacture a comparison.
3Component splits carry the most information per character
Two labelled numbers name the architecture while stating the size, so a reader who recognises the labels knows roughly how the run will behave before touching it. A total would be arithmetically true and less useful.
It also shows which component a memory figure is dominated by, which is the practical question a count is usually asked in place of.
- Model sizeNumber of model parameters: 2,059,174,912published as an exact count
- Model sizeOver 13 billion parametersstated in the repository
- Model sizeVAE: 175M, DiT: 2.8Bpublished as two counts rather than one total
- Model sizeA 10 billion parameter diffusion model, with a 362M parameter video autoencoderstated in the repository
- Model size30 billion parametersfrom the repository's own introduction
4Sources
Repository statements and hosted rates are linked from the model pages and the how read page, checked 2026-09-12. Hardware figures are set out on requirements. Nearby: Complete output, Two figures, one run, Which variant.