Cosmos-Predict2.5-2B
NVIDIA
Cosmos-Predict2.5-2B is a 2,059,174,912-parameter diffusion transformer in NVIDIA's Cosmos world foundation model family, aimed at physical AI such as robotics and autonomous vehicles. It unifies Text2World, Image2World and Video2World in one model and uses Cosmos-Reason1, a physical-AI reasoning VLM, as its text encoder. From a text prompt under 300 words plus an image or a 5-frame video it generates a 5-second 1280x704 mp4 at 16 FPS, requiring 32.54 GB of GPU VRAM. Weights are under the NVIDIA Open Model License; source code is Apache 2.0.
Strengths
- Unifies Text2World, Image2World and Video2World in a single 2B-parameter model
- Uses Cosmos-Reason1, a physical-AI reasoning vision language model, as the text encoder rather than a generic text encoder
- Generates 1280x704 (720P) clips at 16 FPS from text plus an image or a 5-frame video
- Fits in 32.54 GB of VRAM, with inference tested on H100, A100 and GB200 across Ampere, Hopper and Blackwell
- Weights are commercially usable, derivative models are permitted, and NVIDIA claims no ownership of generated outputs
- Source code released under Apache 2.0, with post-training recipes for LoRA, robot action-conditioned and multiview variants
- Later vendor additions include Diffusers support and a distilled Text2World checkpoint
Caveats
- Superseded: NVIDIA states the repository 'is no longer under active development and will receive only limited maintenance updates', with future model releases focused on Cosmos 3, which it says delivers significantly stronger world prediction.
- Output is capped at a 5-second, 16 FPS clip, and NVIDIA warns the model can 'struggle to generate long, high-resolution videos without artifacts', with temporal inconsistency and camera and object motion instability.
- Physics is not guaranteed despite the physical-AI framing: the card notes imprecise interactions and inaccurate representation of physical laws.
- Open-weight, not open-source: the weights carry the NVIDIA Open Model License rather than an OSI licence, and the Hugging Face repository is gated so downloads require accepting that agreement.
- Heavy and slow to run - 32.54 GB VRAM, with inference time from roughly 124 seconds on B200 to about 2,567 seconds on L40S - and the model card publishes no benchmark scores as text.
- Vendor
- NVIDIA
- License
- NVIDIA Open Model License
- Release Date
- 2025-10-06
- Modalities
- textimagevideo
Capabilities
Resources
Pricing
Open weights with no vendor list price. Download is gated on Hugging Face (accepting the NVIDIA Open Model License), and self-hosting needs 32.54 GB of GPU VRAM on Ampere, Hopper or Blackwell hardware. A custom licence is available by contacting [email protected].