Published families that no hosted service names
HunyuanVideo publishes over 13 billion parameters and Mochi 1 publishes 10 billion plus a 362M autoencoder. Neither is named on any hosted service's priced page in this register, so neither has a per-second rate to be set against. As of 2026-09-12.
| Item | What is published | Detail |
|---|---|---|
| HunyuanVideo | Over 13B parameters | Not named by any hosted service here |
| Mochi 1 | 10B plus a 362M autoencoder | Not named by any hosted service here |
| Wan, LTX-Video | Have hosted counterparts | Sold by the second under a later version |
Inclusion rule. A counterpart exists when a hosted service names the family on a page carrying a price. A service that sells unnamed tiers cannot be matched to any family. Order. Families without a counterpart first.
1A cost comparison needs two sides and these have one
For the families with a hosted counterpart, a reader can weigh hardware against a per-second rate, with the caveat that the versions differ. For these two there is no rate at all, so the only cost visible is the hardware one.
That is not a criticism of either family. It is a statement about what this register can compute, and stating it stops a reader looking for a break-even figure that does not exist.
2Parameter counts are the field most often quoted and least often useful
Thirteen billion against ten billion tells a reader almost nothing about output, cost or hardware. It is recorded because repositories publish it and readers look for it, not because it predicts anything the other fields do not say better.
The memory figures and the named cards are the fields that decide whether a model can be run; the parameter count mostly decides how the model is described in headlines.
3A family can gain a counterpart at any time
The moment a service names one of these on a priced page, the row changes and a comparison becomes possible. That is the ordinary way this register moves: nothing about the weights changes, and what a buyer can work out does.
Until then the honest description is published, downloadable, and not for sale by the second anywhere this register reads.
- Model sizeOver 13 billion parametersstated in the repository
- Model sizeA 10 billion parameter diffusion model, with a 362M parameter video autoencoderstated in the repository
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: Licence vs weights, Which version, Running it locally.