NodeValid

Published video-model weights, licences and hardware notes

Allegro: 9.3G with offload, 27.5G without

The model card states 9.3G in BF16 with cpu_offload enabled and 27.5G for the same run with the offload call left out. Both figures describe the same output: 720 x 1280, 88 frames, six seconds at 15 FPS. As of 2026-09-22.

Allegro in the hardware column, as publishedThe memory figure sits on top of a configuration, and the configuration sits on top of an output size and a clip length. Removing any layer changes the number above it.The memory statement as published9.3G BF16 with cpu_offload27.5G with the offload call left out, same run720 x 1280From the model info table on the card88 frames, 6 seconds at 15FPSFrames, seconds and frame rate all statedBoth figures describe the same six seconds
Fig. 1 Two numbers from one card, stacked on the settings that produced them.
Allegro in the hardware column, with the columns that qualify it. Read from huggingface.co/rhymes-ai on 2026-09-22.
ColumnWhat the pages state
Hardware, in the vendor's own words9.3G BF16 with cpu_offload
The detail published beside it27.5G with the offload call left out, same run
Resolution, from the same pages720 x 1280
Length, from the same pages88 frames, 6 seconds at 15 FPS

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.

1Publishing both numbers is the useful part

Offloading moves weights out to system memory between steps, so it buys a third of the requirement back and pays for it in wall-clock time. A card that printed only the 9.3G figure would be accurate and would mislead anyone budgeting for throughput.

Printing both turns one claim into a choice. A reader with a 12G card knows it fits with offload; a reader with a 32G card knows it fits without, and that the run will be faster.

2The output the figures belong to is stated too

Memory scales with pixels and with frames, and both are pinned here: 720 by 1280, 88 frames, six seconds at 15 FPS. That makes the pair of figures transferable, because a reader planning shorter clips knows which direction the number moves.

Most cards in this register attach a memory figure to a resolution and stop. Attaching a frame rate as well is rarer, and it is the difference between 88 frames meaning six seconds and 88 frames meaning something the reader has to work out.

3What is still not settled

Neither figure is a floor for a card. It is the memory a particular script consumed on whatever GPU the vendor used, and a card with the same capacity but an older architecture may not support BF16 at all.

Nor is the trade linear for anybody else. How much time offloading costs depends on host memory bandwidth, which is a property of the machine rather than of the release.

4Sources

Every figure above is the vendor's own, taken from huggingface.co/rhymes-ai on 2026-09-22 with nothing measured here. The rest of the release is on its family entry, the whole column is on hardware, and all memory figures together are on requirements. Nearby: Pyramid Flow, no count, CogVideoX split licence.