Counted to the unit against over thirteen billion
Cosmos Predict 2.5 states the number of model parameters as 2,059,174,912. HunyuanVideo states over 13 billion parameters. One is a measurement of an artefact, the other is a bound with no upper end. As of 2026-09-22.
| Point | Cosmos Predict 2.5 | HunyuanVideo |
|---|---|---|
| Size, as each vendor states it | 2,059,174,912 parameters | Over 13 billion parameters |
| The detail published beside it | Published as an exact count, not a rounded figure | Stated in the repository |
| Licence, from the same pages | NVIDIA Open Model License | tencent-hunyuan-community |
| Hardware, from the same pages | 32.54 GB of GPU VRAM | 60GB peak at 720px1280px, 129 frames |
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.
1Why a hedge is a defensible choice
Counting conventions differ: text encoders in or out, autoencoders separate, embeddings sometimes ignored. Writing over thirteen billion avoids committing to a convention the next paper may not share.
It is reasonable for an author and inconvenient for a register, which holds this column specifically so that two rows can be placed beside each other.
2What exactness actually buys
It removes the convention question entirely. A count to the unit is a statement about one file, so nobody has to ask what was included, and the same card gives a memory figure to two decimals.
Precision does not imply correctness or breadth. Both figures still describe one release at one point in time, and the register keeps the date it read them.
3The bound is consistent with the rest of its page
Over thirteen billion parameters sits beside 60GB peak at 720px1280px and 45GB at 544px960px, both on a single 80G card. Nothing about that page contradicts the bound.
So the hedge is not concealment. It is a release documented by measurements of runs rather than by measurements of files.
- Model sizeNumber of model parameters: 2,059,174,912published as an exact count
- Model sizeOver 13 billion parametersstated in the repository
4Sources
Read from huggingface.co/nvidia and github.com/Tencent-Hunyuan on 2026-09-22. Full entries: Cosmos Predict 2.5 and HunyuanVideo. What this column is quoting is on size. Nearby pairs: Decomposed or round, No count, or three.