Allegro: what the repository publishes

Allegro publishes a 175M autoencoder beside a 2.8B transformer under the apache-2.0 tag, and states single-GPU memory twice: 9.3G with CPU offload switched on, 27.5G with it off. Output is 720 x 1280, 88 frames, 6 seconds at 15 FPS. As of 2026-09-22.

Inclusion rule. The same five fields for every family here. A field the repository does not state is recorded as not stated rather than estimated from the parameter count. Order. Fixed field order, identical on every family page.

1Two memory figures for one run, and the switch between them is named

The card does not publish a memory requirement. It publishes a memory usage: 9.3G in BF16 with CPU offload enabled, and 27.5G if the offload call is left out. Both describe the same generation, and the card names the line of code that moves between them while warning that inference time will increase significantly on the cheaper path.

That shape is worth more than a single floor. Somebody holding a 12GB card learns the run is possible and slow; somebody holding a 32GB card learns the fast path fits. A card that published only the low figure would have implied the fast path fits everywhere, and one that published only the high figure would have written off hardware that works.

2A parameter count in two parts, because the download is in two parts

The card gives VAE: 175M and DiT: 2.8B rather than a total. The autoencoder and the diffusion transformer are separate files doing separate jobs, and adding them into 2.975B would produce a figure the vendor never published and which predicts nothing about either piece.

This register keeps both because the card keeps both. One other entry here splits its count the same way, and in both cases the split carries more information: the transformer is what grows between model generations while the autoencoder tends to stay roughly where it was.

3Length stated three ways at once

Frames, seconds and frame rate together: 88, six seconds, 15 FPS. Most repositories here publish one of those three and leave a reader to reconstruct the others, which cannot be done without the frame rate. This row needs no arithmetic and carries no caveat.

The 15 FPS is part of the claim rather than a footnote to it. It sits below what most delivery formats assume, so six seconds is six seconds of 15 FPS material, and anything that has to cut against 24 or 30 FPS footage acquires a conversion step the card never discusses.

4The same field, across every family

Inclusion rule. Families whose vendor has published weights openly. A family whose repository does not state this particular field keeps its row, marked hollow. Order. Alphabetical by family name.

Inclusion rule. Families whose vendor has published weights openly. A family whose repository does not state this particular field keeps its row, marked hollow. Order. Alphabetical by family name.

Inclusion rule. Families whose vendor has published weights openly. A family whose repository does not state this particular field keeps its row, marked hollow. Order. Alphabetical by family name.

• Licence
  apache-2.0the tag on the vendor's model cardrhymes-ai/Allegro on Hugging Face / recorded 2026-09-22
• Model size
  VAE: 175M, DiT: 2.8Bpublished as two counts rather than one totalrhymes-ai/Allegro on Hugging Face / recorded 2026-09-22
• Hardware
  9.3G BF16 with cpu_offload, against 27.5G if CPU offload is not enabledsingle GPU memory usage on the same runrhymes-ai/Allegro on Hugging Face / recorded 2026-09-22
• Resolution and length
  720 x 1280, 88 frames, 6 seconds at 15 FPSfrom the model info tablerhymes-ai/Allegro on Hugging Face / recorded 2026-09-22

5Sources

Read from huggingface.co/rhymes-ai on 2026-09-22. Hardware figures across every family are collected on requirements; the licence column is set out on licence and what counts as a published weight is on how read. Previous family: Mochi 1. Next family: Step-Video-T2V.