NodeValid

Published video-model weights, licences and hardware notes

Cosmos states 32.54 GB and names the chips

The card states that the model requires 32.54 GB of GPU VRAM and names Ampere, Hopper and Blackwell as supported hardware. The output it buys is 720P at 16FPS, as a five-second clip. As of 2026-09-22.

Cosmos Predict 2.5 in the hardware column, as publishedNaming architectures rather than card models is a more durable statement, because new cards join an architecture while a model number ages immediately.The memory statement as published32.54 GB of GPU VRAMAmpere, Hopper and Blackwell named as supported720PStated with a 16FPS frame rateA 5-second clipSeconds, stated alongside the frame rateArchitectures outlast model numbers
Fig. 1 An exact figure plus the chip generations it applies to, rather than a card list.
Cosmos Predict 2.5 in the hardware column, with the columns that qualify it. Read from huggingface.co/nvidia on 2026-09-22.
ColumnWhat the pages state
Hardware, in the vendor's own words32.54 GB of GPU VRAM
The detail published beside itAmpere, Hopper and Blackwell named as supported
Resolution, from the same pages720P
Length, from the same pagesA 5-second clip

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.

1Architectures age better than card names

A repository that names an RTX 4090 has to be edited when the next generation ships. One that names Ampere, Hopper and Blackwell covers cards that did not exist when the card was written.

It also tells a reader what the constraint actually is, which is the numeric formats and the tensor core generation rather than the badge on the box.

2Thirty-two gigabytes for two billion parameters

The count on this card is 2,059,174,912, and the memory figure is over thirty gigabytes. Weights in half precision account for a small share of that; the rest is the run.

The other 2B release in the register states no memory figure at all, so these two rows together bracket what a reader can infer from a count, which is nothing.

3A complete output specification, for once

720P, 16FPS and five seconds are all on the same card, so the memory figure has a fully described output attached to it. Most rows in the register leave at least one of the three out.

That completeness is what makes the figure transferable. A reader planning ten seconds knows exactly which part of the specification they are changing.

4Sources

Every figure above is the vendor's own, taken from huggingface.co/nvidia 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: No card, no figure, A 4090 and 80GB.