whatmodel.sh / video / Cosmos-Predict2.5-2B
Cosmos-Predict2.5-2B
↗ share
+ contribute
Cosmos-Predict2.5-2B is an open-weights video model from NVIDIA, released October 2025. Its published size is 2B parameters. Licence: nvidia-open-model-license.
Catalogue identity notes
NVIDIA Cosmos-Predict2.5 2B world foundation model (one checkpoint covering Text2World / Image2World / Video2World), released 6 Oct 2025 (cosmos-predict2.5 README news). HF repo nvidia/Cosmos-Predict2.5-2B pinned 85f8ae7 (gated, automatic approval; licence 'other' = NVIDIA Open Model License). PAI-Bench Table 12 pins the checkpoint as Version 'base/post-trained' (the repo's base/post-trained folder) at 93 frames, 16 fps, 1280x704; PhyGround Appendix lists 'Cosmos-Predict2.5-2B' (93 frames, 16 fps, 1280x704, 36 steps) as its 'Cosmos-2B' row; the Cosmos 3 report (NVIDIA) runs it as 'Cosmos-Predict2.5-2B' on Cosmos HUE. Comparison-only: no local command was verified in this pass (the official route is the cosmos-predict2.5 repo or the diffusers Cosmos2_5_PredictBasePipeline). Distinct from cosmos3-nano / cosmos3-super and the older cosmos-1-0-* entries. Not mapped: Rapidata 'Cosmos Predict 2.5' rows (size not stated).
Model information sources
Release date ↗ · Weights licence ↗
2B as named by NVIDIA (the size in the model name); the HF repo is gated (automatic approval), so no safetensors count was read. Text encoder and tokenizer not included.
$ how it compares
The video models ranked either side of Cosmos-Predict2.5-2B, out of 139 with a score, then more from NVIDIA.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 72 | Cosmos-1.0-Diffusion-7B-Video2World | 49±5 | comparison only | ||
| 73 |  Seedance 1.0 Lite | 48±2 | comparison only | ||
| 74 | Cosmos-Predict2.5-2B | 48±2 | comparison only | ||
| 75 | Cosmos-Predict2-14B-Video2World | 48±8 | comparison only | ||
| 76 |  MAGI-1 24B | 48±4 | comparison only | ||
| more from NVIDIA | |||||
| 38 | Cosmos3-Super | 58±4 | local | ||
| 60 | Cosmos3-Nano | 52±3 | local | ||
| 79 | Cosmos-Predict2-2B-Video2World | 48±8 | comparison only | ||
$ what it's good at
text to video▇▅▄▆█▆▇▂▃▃▂▄ #51 of 94 * · 22% of its score · 1 test
Trails Hailuo 01 (MiniMax video-01 / I2V-01), Cosmos-1.0-Diffusion-7B-Text2World, LTX-2 19B Distilled here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Cosmos HUE human evaluation, text-to-video (Cosmos 3 report) lab | 81.8% ✎ | █▄▄··▄▄▄▄▄·▄ beats 1/9 | 22% |
image to video▁▁▂▂██▄▃▃▄▂▂ #50 of 74 · 69% of its score · 4 tests
Beats HunyuanVideo 1.5, Wan 2.2 A14B, LTX-2 here, though they rank higher overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| CRONOS counterfactual rigid-body physics (I2V success rate) | 12.0% ✎ | █·····█··█·█ beats 2/3 | 24% |
| PAI-Bench-G human preference arena (Overall ELO) | 1008.1 ✎ | ▄······██··█ beats 4/6 | 21% |
| Cosmos HUE human evaluation, image-to-video (Cosmos 3 report) lab | 82.6% ✎ | ▄·█··▄··▄▄▄█ beats 1/8 | 17% |
| PAI-Bench-G physical-AI video generation (Overall) | 81.4% ✎ | ▃······▅·█▅█ beats 8/10 | 7.0% |
other tests1 not in a task · 6.3% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| PhyGround physics human evaluation (Overall, TI2V) | 51.6% ✎ | █▄▄··▄···▄·█ beats 1/7 | 6.3% |
$ where to run it
No verified API or download recorded yet · add one
Official weights and documentation: huggingface.co/nvidia ↗
$ how the score was worked out · Cosmos 3: Omnimodal World Mod… 39%, PAI-Bench: A Comprehensive Be… 28%, CRONOS: Benchmarking Counterf… 24% and 1 more
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 Cosmos 3: Omnimodal World Models for Physical AI (arXiv:2606.02800v4) · lab-reported | 2 | 39% |
| PAI-Bench: A Comprehensive Benchmark For Physical AI (arXiv:2512.01989v1) | 2 | 28% |
| CRONOS: Benchmarking Counterfactual Physical Consistency in Video Models (arXiv:2605.23699v1), Table 1 | 1 | 24% |
| PhyGround: Benchmarking Physical Reasoning in Generative World Models (arXiv:2605.10806v1) | 1 | 6% |
| prior · pull toward the middle | 3% | |
| weighted average | = 48.4 |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json