NodeValid

Published video-model weights, licences and hardware notes

Mochi counts the autoencoder separately

The repository states a 10 billion parameter diffusion model and, separately, a 362M parameter video autoencoder. Two numbers for one download, which is a more honest shape than a single total would be. As of 2026-09-12.

Mochi 1 in the size column, as publishedA video model is several trained components, and adding their counts together hides which of them a memory figure or a licence actually attaches to.Fixed by the published countLeft to the runThe count10 billion parameters—What it belongs toPlus a 362M video autoencoder—Licence on that releaseApache 2.0—Memory for a run—About 60GB VRAM on a single GPUOutput size—480pA total would have lost the decomposition
Fig. 1 Two components, two counts, and one number that would have concealed both.
Mochi 1 in the size column, with the columns that qualify it. Read from github.com/genmoai on 2026-09-12.
ColumnWhat the pages state
Size, in the vendor's own words10 billion parameters
The detail published beside itPlus a 362M video autoencoder
Licence, from the same pagesApache 2.0
Hardware, from the same pagesAbout 60GB VRAM on a single GPU

Inclusion rule. The value this page is about, the detail the vendor attached to it, and the two columns that change how it should be read. A column the vendor left empty keeps its row and says so. Order. This column first, then its detail, then the two columns that qualify it.

1Why a total would be worse

The autoencoder and the diffusion model do different jobs and are loaded at different points in a run. A combined figure of about 10.4 billion would be arithmetically true and would tell a reader less than the pair does.

It also matters for anyone replacing one component. Decoders get swapped, quantised and reused; a release that names its parts makes that a visible option.

2The memory figure belongs to the whole run

About 60GB on a single GPU, with at least one H100 recommended, covers both components plus activations. Neither count predicts it, and the ratio between the counts does not predict how the memory splits.

This is the general reason size and hardware are separate columns here, and this row is a clean illustration: the smaller component is under four per cent of the parameters.

3What the row leaves open

Nothing in the repository says which component the 480p limit comes from, and the resolution is described as the current capability rather than as an architectural ceiling.

So the counts describe the files, and the output description describes the release as it stood when the page was read. The register keeps both and joins neither.

  • Model size
    A 10 billion parameter diffusion model, with a 362M parameter video autoencoderstated in the repositoryGenmo, Mochi repository / recorded 2026-09-12
  • Hardware
    Approximately 60GB VRAM on a single GPU, with at least one H100 recommendedthe vendor's own figureGenmo, Mochi repository / recorded 2026-09-12

4Sources

The count comes from github.com/genmoai, read 2026-09-12, exactly as the vendor wrote it. Everything else this release publishes is on its family entry; why the count sits in its own column is on size. Nearby: VAE and DiT apart, Thirty billion.