NodeValid

Published video-model weights, licences and hardware notes

Allegro publishes VAE and DiT separately

The model info table gives two counts: VAE at 175M and DiT at 2.8B. The split names the architecture while stating the size, and it makes clear which of the two the memory figures are dominated by. As of 2026-09-22.

Allegro in the size column, as publishedNaming the components as VAE and DiT tells a reader what kind of model this is before they have read a single line of the paper.Fixed by the published countLeft to the runThe countVAE 175M and DiT 2.8B—What it belongs toPublished as two counts, not one total—Licence on that releaseapache-2.0—Memory for a run—9.3G BF16 with cpu_offloadOutput size—720 x 1280The split names the design
Fig. 1 A size row that doubles as an architecture note, at no extra length.
Allegro in the size column, with the columns that qualify it. Read from huggingface.co/rhymes-ai on 2026-09-22.
ColumnWhat the pages state
Size, in the vendor's own wordsVAE 175M and DiT 2.8B
The detail published beside itPublished as two counts, not one total
Licence, from the same pagesapache-2.0
Hardware, from the same pages9.3G BF16 with cpu_offload

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.

1The labels carry information the numbers do not

VAE and DiT are specific: a variational autoencoder for the latent space and a diffusion transformer for the denoising. A reader who knows those two words knows roughly how the run will behave before touching it.

Compare a card that says 3B and nothing else. Same order of magnitude, far less to go on, and no way to guess which part quantising would help.

2Which component the memory belongs to

At 9.3G with cpu_offload and 27.5G without, the swing is far larger than the 175M autoencoder could account for. The offload is moving the transformer, and the split makes that legible.

That is the practical value of decomposing a count. It turns a memory figure from a fact to be memorised into something with a cause.

3Under three billion, and documented like a bigger release

At 2.8B for the main component this is one of the smaller releases in the register, yet the card publishes a frame rate, a frame count, a resolution and two memory figures.

Documentation quality does not track model size. Some of the largest releases here state less about their outputs than this card does.

4Sources

The count comes from huggingface.co/rhymes-ai, read 2026-09-22, 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: Thirty billion, Size in the release name.