Cosmos Predict 2.5: what the repository publishes

Cosmos Predict 2.5 fills every cell: the NVIDIA Open Model License with a sentence saying models are commercially usable, 2,059,174,912 parameters, a requirement of 32.54 GB of GPU VRAM, and 720P video at 16FPS delivered as a 5-second clip. As of 2026-09-22.

Inclusion rule. The same five fields for every family here. A field the repository does not state is recorded as not stated rather than estimated from the parameter count. Order. Fixed field order, identical on every family page.

1The fullest row here, and the precision is the interesting part

A parameter count to the digit, a memory figure to two decimal places, a frame rate, a clip length and a licence that addresses commercial use inside the file. Every other entry in this register leaves at least one of those for the reader to find elsewhere or to guess.

Precision is not accuracy, and the card is careful about which it is offering: the VRAM figure is written as a requirement for this model rather than as a measurement of one particular run. A reader still has to ask what settings it assumes. They do not have to ask what the number is.

2A licence that says commercially usable inside its own text

The card states the model is released under the NVIDIA Open Model License and that models are commercially usable. That is a vendor-written licence answering, in its own words, the question most licences in this register hand off to a separate page or leave alone.

The register still records the name rather than the permission, for the same reason it does everywhere else: a sentence on a model card summarises a document, and the document governs. But the presence of that sentence is itself a fact, and it is the reason this row reads differently from the other vendor-written licences here.

3Architectures named, not just one card

Ampere, Hopper and Blackwell are listed as supported, which is a statement about card generations rather than about a single product. Anybody holding a card can place it in a generation without doing any memory arithmetic first.

Naming generations and stating 32.54 GB together is what makes a hardware plan checkable from published material: the generation settles whether the numeric formats are supported, the figure settles whether the memory is. Most rows in this register carry one of those two and leave the other open.

4The same field, across every family

Inclusion rule. Families whose vendor has published weights openly. A family whose repository does not state this particular field keeps its row, marked hollow. Order. Alphabetical by family name.

Inclusion rule. Families whose vendor has published weights openly. A family whose repository does not state this particular field keeps its row, marked hollow. Order. Alphabetical by family name.

Inclusion rule. Families whose vendor has published weights openly. A family whose repository does not state this particular field keeps its row, marked hollow. Order. Alphabetical by family name.

• Licence
  NVIDIA Open Model License, with the card stating that models are commercially usablea vendor-written licence answering the commercial question in its own textnvidia/Cosmos-Predict2.5-2B on Hugging Face / recorded 2026-09-22
• Model size
  Number of model parameters: 2,059,174,912published as an exact countnvidia/Cosmos-Predict2.5-2B on Hugging Face / recorded 2026-09-22
• Hardware
  This model requires 32.54 GB of GPU VRAM, on Ampere, Hopper or Blackwell hardwarea figure and the supported architecturesnvidia/Cosmos-Predict2.5-2B on Hugging Face / recorded 2026-09-22
• Resolution and length
  Produces 720P video with 16FPS, as a 5-second clipboth stated on the cardnvidia/Cosmos-Predict2.5-2B on Hugging Face / recorded 2026-09-22

5Sources

Read from huggingface.co/nvidia on 2026-09-22. Hardware figures across every family are collected on requirements; the licence column is set out on licence and what counts as a published weight is on how read. Previous family: MAGI-1. Next family: LongCat-Video.