NodeValid

Published video-model weights, licences and hardware notes

Which settings are published beside a memory figure

Across the register, the output resolution is the setting most often published beside a memory figure. The numeric precision appears less often, the frame count less often again, and the batch size once. As of 2026-09-12.

Four dials behind every published figurePrecision, resolution, frame count and offloading each move a memory figure, and a statement that names none of them cannot be adjusted by a reader.Named most often beside a figureResolutionSix families state the output size the figure was measured at.PrecisionFour state the numeric format the figure assumes.Frame countThree state how many frames were held at once.Batch sizeOne states it, which removes the largest hidden variable.Named dials are what make a figure reusable
Fig. 1 How often each dial is actually named beside a number in this register.
Settings published beside memory figures, counted across the register. Recorded 2026-09-12.
ItemWhat is publishedDetail
Output resolutionSix families720px1280px, 768px768px, 768px, 540P, 720P and 720 x 1280
Numeric precisionFour familiesFP16, BF16 twice, and an fp8 build
Frame countThree families129 frames, 204 and 136 frames, and 1800 frames
Offload stateOne familyWith cpu_offload enabled, and with the call left out
Batch sizeOne familyBatch size one, stated for three peak figures
Card or architectureEight familiesA card model, a card size or a chip generation

Inclusion rule. One row per setting that changes a memory figure. A setting stated by one family keeps its row, because rarity is the finding. Order. From the setting stated most often to the settings stated once.

1Batch size is the largest hidden variable

Weights are held once regardless of batch size and everything else scales with it, so an unstated batch makes a figure ambiguous in one direction. A figure at batch size four is a much lower per-stream cost than it looks.

One repository here states it, which means those figures are the lowest the vendor could honestly publish. Everywhere else a reader has to assume one and know they are assuming.

2Resolution is stated most often because it is easiest

A memory table naturally has rows, and the rows are usually resolutions. That is why six families name an output size beside a figure and only one names a batch size: the first falls out of the table's shape and the second has to be written deliberately.

It is also the setting readers most want, since resolution is the lever they are most likely to pull. So the column is well served by accident on its most useful axis.

3A figure with no settings cannot be adjusted

Four families publish gigabytes with no hardware named, and some of those also omit the precision. A reader wanting a different resolution or a shorter clip has no published starting point to move from.

The register keeps each figure with whatever settings were published and adds none. A configuration supplied by the register would be the part a reader relied on and the part nobody stated.

  • Hardware
    78.55 GB peak GPU memory at 768px768px204f, 77.64 GB at 544px992px204f and 72.48 GB at 544px992px136fstated per configuration at batch size onestepfun-ai, Step-Video-T2V repository / recorded 2026-09-22
  • Hardware
    From 4GB in FP16 for the 2B model, from 5GB in BF16 for the 5B, and from 10GB for CogVideoX1.5-5Bthe lowest figures the repository statesTHUDM, CogVideo repository / recorded 2026-09-12
  • 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
  • Hardware
    Peak GPU memory of 60GB at 720px1280px and 129 frames, and 45GB at 544px960pxthe vendor's own figuresTencent-Hunyuan, HunyuanVideo repository / recorded 2026-09-12
  • Hardware
    For 4.5B models, any machine with at least 24GB of GPU memory is sufficient, and a distilled fp8 configuration works on GPUs with at least 12GBa floor for the model and a lower one for a buildSandAI-org, MAGI-1 repository / recorded 2026-09-22

4Sources

Repository statements and hosted rates are linked from the model pages and the how read page, checked 2026-09-12. Hardware figures are set out on requirements. Nearby: Card or repository, More than one card, Empty cells.