NodeValid

Published video-model weights, licences and hardware notes

CogVideoX starts at 4GB for its 2B model

The repository states from 4GB in FP16 for the 2B model, from 5GB in BF16 for the 5B, and from 10GB for CogVideoX1.5-5B. Three figures, each tied to a named release and a named precision. As of 2026-09-12.

CogVideoX in the hardware column, as publishedA ladder of three figures inside one family, each with the precision it assumes, is the most transferable hardware statement in the register.The memory statement as publishedFrom 4GB in FP16 on the 2BFrom 5GB on the 5B, from 10GB on 1.5-5B1360x768 on 1.5-5B720x480 on the 5BUp to 161 frames on 1.5-5BDescribed as 5 or 10 secondsThe word from is doing real work
Fig. 1 Three releases, three floors, and the precision written next to each one.
CogVideoX in the hardware column, with the columns that qualify it. Read from github.com/THUDM on 2026-09-12.
ColumnWhat the pages state
Hardware, in the vendor's own wordsFrom 4GB in FP16 on the 2B
The detail published beside itFrom 5GB on the 5B, from 10GB on 1.5-5B
Resolution, from the same pages1360x768 on 1.5-5B
Length, from the same pagesUp to 161 frames on 1.5-5B

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.

1Why this row is unusually usable

The figures are written as floors rather than as peaks, each is attached to a named release, and each names the numeric format it assumes. A reader can match their own card against it without guessing at any of the three.

Very few rows in the register offer all three. Most give a peak, or a figure without a precision, or a figure without saying which variant it belongs to.

2Four gigabytes buys the smallest of three models

The 4GB figure belongs to the 2B release, which is also the one under standard licence terms. It is not the figure for the model that produces 1360x768 at up to 161 frames.

That release is the 1.5-5B, and it starts at 10GB. Reading the family's lowest figure as the family's requirement would be wrong by a factor of two and a half.

3What the ladder still assumes

That the card supports the precision named. A 4GB card that cannot run FP16 does not qualify, and the word from implies a configuration the repository chose rather than the one a reader will use.

Longer clips and higher resolutions move all three figures upward. The ladder is a starting point, published at the bottom of its own range.

  • 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
  • Length and resolution
    CogVideoX1.5-5B states 1360x768 and up to 161 frames, described as 5 or 10 secondsper modelTHUDM, CogVideo repository / recorded 2026-09-12

4Sources

Every figure above is the vendor's own, taken from github.com/THUDM on 2026-09-12 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: Sixty gigabytes and an H100, Never below 72GB.