NodeValid

Published video-model weights, licences and hardware notes

From four gigabytes, or about sixty

CogVideoX states from 4GB in FP16 for its 2B release. Mochi 1 states approximately 60GB on a single GPU with at least one H100 recommended. Both are single-card figures and they are fifteen times apart. As of 2026-09-12.

CogVideoX and Mochi 1 on the same columnTwo single-card floors fifteen times apart, one for a small model at low precision and one for a larger model at standard definition output.CogVideoXMochi 1Hardware, as each vendor statesitFrom 4GB in FP16 on the 2BAbout 60GB VRAM on a single GPUThe detail published beside itFrom 5GB on the 5B, from 10GB on 1.5-5BAt least one H100 recommendedResolution, from the same pages1360x768 on 1.5-5B480pBoth are floors, on one card each
Fig. 1 The bottom and the upper end of the single-card range, as published.
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 stateCogVideoXMochi 1Hardware, as each vendor states…Hardware, as each vendor states itHardware, as each vendor states it — CogVideoX: From 4GB in FP16 on the 2BHardware, as each vendor states it — Mochi 1: About 60GB VRAM on a single GPUThe detail published beside itThe detail published beside it — CogVideoX: From 5GB on the 5B, from 10GB on 1.5-5BThe detail published beside it — Mochi 1: At least one H100 recommendedResolution, from the same pagesResolution, from the same pages — CogVideoX: 1360x768 on 1.5-5BResolution, from the same pages — Mochi 1: 480pLength, from the same pagesLength, from the same pages — CogVideoX: Up to 161 frames on 1.5-5BLength, from the same pages — Mochi 1: No length recorded
Fig. 2 Filled where the repository states something, hollow where it does not. A blank is not an inference.
CogVideoX and Mochi 1 on the hardware column, read from github.com/THUDM and github.com/genmoai. Recorded 2026-09-12.
PointCogVideoXMochi 1
Hardware, as each vendor states itFrom 4GB in FP16 on the 2BAbout 60GB VRAM on a single GPU
The detail published beside itFrom 5GB on the 5B, from 10GB on 1.5-5BAt least one H100 recommended
Resolution, from the same pages1360x768 on 1.5-5B480p
Length, from the same pagesUp to 161 frames on 1.5-5BNo length 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.

1This pair is comparable in the way most are not

Both figures are floors rather than peaks, both are for a single GPU, and both name the release they belong to. That makes the fifteenfold gap a difference between models rather than between kinds of statement.

The precision is stated on one side and not the other, which is the one asymmetry left. FP16 is named for the 4GB figure; Mochi's approximately 60GB has no format attached.

2The output sizes run the other way

CogVideoX's larger release produces 1360x768 with up to 161 frames. Mochi produces 480p, described as the current capability. The release with the higher memory figure has the lower stated output.

That is a reminder that memory tracks architecture and training rather than resolution, and that the register's columns cannot be read as a single ranking.

3Which figure to plan against

Neither, without reading the qualifier. CogVideoX's 4GB belongs to the 2B release, not to the 1360x768 one, which starts at 10GB. Mochi's 60GB comes with a card recommendation that a memory total does not satisfy.

The two rows together are the argument for keeping the setting attached to every figure in this column.

  • 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
    Approximately 60GB VRAM on a single GPU, with at least one H100 recommendedthe vendor's own figureGenmo, Mochi repository / recorded 2026-09-12

4Sources

Read from github.com/THUDM and github.com/genmoai on 2026-09-12. Full entries: CogVideoX and Mochi 1. What this column is quoting is on hardware. Nearby pairs: Cards or gigabytes, Two pairs, two variables.