NodeValid

Published video-model weights, licences and hardware notes

A count split in two against one round figure

Allegro publishes VAE at 175M and DiT at 2.8B. Step-Video-T2V introduces itself as a 30 billion parameter model. One tells a reader what the model is made of; the other tells them how large it is. As of 2026-09-22.

Allegro and Step-Video-T2V on the same columnDecomposing a count names the architecture at the same time as the size, which a single round figure cannot do at any level of precision.AllegroStep-Video-T2VSize, as each vendor states itVAE 175M and DiT 2.8B30 billion parametersThe detail published beside itPublished as two counts, not one totalFrom the repository's own introductionLicence, from the same pagesapache-2.0MITComponents, or a single number
Fig. 1 Two counts that describe a design, against one that describes a magnitude.
Allegro and Step-Video-T2V on the size column, read from huggingface.co/rhymes-ai and github.com/stepfun-ai. Recorded 2026-09-22.
PointAllegroStep-Video-T2V
Size, as each vendor states itVAE 175M and DiT 2.8B30 billion parameters
The detail published beside itPublished as two counts, not one totalFrom the repository's own introduction
Licence, from the same pagesapache-2.0MIT
Hardware, from the same pages9.3G BF16 with cpu_offload78.55 GB peak at 768px768px204f

Inclusion rule. Two families whose published values in this column lead to opposite decisions. A pair that happens to agree is not listed, because agreement is already visible on the column summary. Order. The column itself first, then its detail, then the columns that qualify it.

1The labels are the information

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

Thirty billion parameters says nothing about structure. It is the right figure for planning storage and the wrong one for guessing which component quantising would help.

2Which component the memory belongs to

Allegro's swing from 9.3G with offload to 27.5G without is far larger than a 175M autoencoder could account for, so the offload is moving the transformer. The split makes that legible.

Step-Video's memory table barely moves across three configurations, from 78.55 GB to 72.48 GB, which says the weights dominate. Different information, arrived at differently.

3Round numbers are descriptions

Thirty billion is round enough to be a characterisation rather than a measurement, and the register does not sharpen it. The repository's own introduction is the source.

Allegro's figures are equally the vendor's own, from the model info table on the card. Neither release is estimated here and neither is combined into a total.

4Sources

Read from huggingface.co/rhymes-ai and github.com/stepfun-ai on 2026-09-22. Full entries: Allegro and Step-Video-T2V. What this column is quoting is on size. Nearby pairs: No count, or three, Count and cost inverted.