NodeValid

Published video-model weights, licences and hardware notes

Publishing both sides of one trade, or none

Allegro publishes 9.3G with cpu_offload and 27.5G without, for one run. LongCat-Video publishes 13.6B parameters, 720p at 30fps and no memory figure of any kind. Four figures against none. As of 2026-09-22.

Allegro and LongCat-Video on the same columnPublishing both sides of the offload trade turns a claim into a choice; publishing neither leaves a reader with nothing to act on.AllegroLongCat-VideoHardware, as each vendor statesit9.3G BF16 with cpu_offloadNo memory figure recordedThe detail published beside it27.5G with the offload call left out,same runThe repository states none and names nocardResolution, from the same pages720 x 1280720pA choice, or an unanswered question
Fig. 1 One card shows the trade, the other does not mention the machine.
What the repositories actually stateFilled where the repository states something, hollow where it does not. A blank is not an inference.What the repositories actually stateAllegroLongCat-VideoHardware, as each vendor states…Hardware, as each vendor states itHardware, as each vendor states it — Allegro: 9.3G BF16 with cpu_offloadHardware, as each vendor states it — LongCat-Video: No memory figure recordedThe detail published beside itThe detail published beside it — Allegro: 27.5G with the offload call left out, same runThe detail published beside it — LongCat-Video: The repository states none and names no cardResolution, from the same pagesResolution, from the same pages — Allegro: 720 x 1280Resolution, from the same pages — LongCat-Video: 720pLength, from the same pagesLength, from the same pages — Allegro: 88 frames, 6 seconds at 15 FPSLength, from the same pages — LongCat-Video: No ceiling recorded
Fig. 2 Filled where the repository states something, hollow where it does not. A blank is not an inference.
Allegro and LongCat-Video on the hardware column, read from huggingface.co/rhymes-ai and github.com/meituan-longcat. Recorded 2026-09-22.
PointAllegroLongCat-Video
Hardware, as each vendor states it9.3G BF16 with cpu_offloadNo memory figure recorded
The detail published beside it27.5G with the offload call left out, same runThe repository states none and names no card
Resolution, from the same pages720 x 1280720p
Length, from the same pages88 frames, 6 seconds at 15 FPSNo ceiling recorded

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.

1Two figures make a lever visible

9.3G against 27.5G for the same output is a third of the memory for more wall-clock time. A reader with a 12G card knows it fits with offload; one with a 32G card knows it fits without.

Neither figure alone would carry that. The friendlier number on its own would mislead anyone budgeting throughput, and the larger one alone would discourage users the release can serve.

2The silent release is the larger one

13.6B against 2.8B for the main Allegro component. The model where memory is more likely to be the obstacle is the one with nothing published about it.

It does publish 720p at 30fps, which is the highest frame rate in the register, and a qualitative length claim of minutes-long. The documentation went elsewhere.

3What the register can and cannot carry across

Allegro's pair is attached to 720 x 1280, 88 frames and six seconds at 15 FPS, so it can be adjusted by a reader who needs a different output.

Nothing about it transfers to a different architecture. The LongCat cell stays empty rather than borrowing a number from a smaller model.

4Sources

Read from huggingface.co/rhymes-ai and github.com/meituan-longcat on 2026-09-22. Full entries: Allegro and LongCat-Video. What this column is quoting is on hardware. Nearby pairs: Bespoke or MIT, Split, or one document.