Cosmos-Predict2.5
NVIDIA, United States 路 October 2025
World model 路 Open weights, NVIDIA Open Model 路 Also written Cosmos Predict 2.5, Cosmos-Predict2.5-2B, Cosmos-Predict2.5-14B
- Parameters
- 2.1 B
- 2,059,174,912 parameters for Cosmos-Predict2.5-2B and 14,368,048,004 for Cosmos-Predict2.5-14B, both from the Hugging Face...
- GPU memory
- 32.54 GB
- reported by the authors
- Inference latency
- not published
- per action step
- Weights
- NVIDIA Open Model
- NVIDIA Open Model License for the checkpoints, which the 14B model card describes as ready for commercial and non-commercial use.
What it is
Cosmos-Predict2.5 is NVIDIA's open video world foundation model for physical AI. One flow-based diffusion transformer handles text to video, image to video and video to video generation, takes its text grounding from the Cosmos-Reason1 vision language model and was trained on 200M curated video clips. NVIDIA positions it for synthetic data generation, policy evaluation and closed-loop simulation, and publishes robot variants: an action-conditioned model that turns an image and a robot action sequence into future frames, multiview robot models and a policy variant post-trained on LIBERO and RoboCasa. On action-conditioned prediction of 100 Bridge test episodes the 2B robot model reaches 24.95 dB PSNR and an FVD of 146, against 21.14 dB and 190 for Cosmos-Predict1-7B. The base model only generates video, and one 5 second 720p clip takes 228.8 s on an H100 SXM.
Architecture
- Backbone
- Diffusion transformer that denoises video in the latent space of the Wan2.1 VAE (4x8x8 compression). Text conditioning comes from the Cosmos-Reason1 vision language model instead of the T5 encoder of earlier versions. The 2B model has 32 layers of width 2,048, the 14B model 36 layers of width 5,120. Developed from Cosmos-Predict2.
- Action head
- None in the base model. The robot action-conditioned variant takes an image and a robot action sequence and predicts the resulting frames at 256p and 4 FPS, with actions injected through the time embedding. A robot policy variant outputs actions, future images and future proprioception.
- Parameters
- 2,059,174,912 parameters for Cosmos-Predict2.5-2B and 14,368,048,004 for Cosmos-Predict2.5-14B, both from the Hugging Face model cards. The action-conditioned, multiview and policy robot variants are listed on the 2B card; the GR00T Dreams robot variant is a 14B model.
- Pretraining data
- 200M curated video clips, followed by reinforcement learning based post-training. The paper's robotics data table lists AgiBot-Beta (194k head-view videos), DROID (39k wrist-view videos) and Bridge (36k videos).
- Embodiments
- WidowX (Bridge videos, action-conditioned variant), AgiBot head and wrist cameras (multiview variant), Fourier GR1 humanoid (GR00T Dreams variant), LIBERO and RoboCasa simulation (policy variant)
What hardware it needs
This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.
| Precision | Weights | Note |
|---|---|---|
| fp32 | 7.7 GB | Training master weights, rarely used for inference |
| bf16 or fp16 | 3.8 GB | The usual way these checkpoints are served |
| int8 | 1.9 GB | Quantised, expect some loss of precision on fine motions |
| int4 | 1.0 GB | Aggressive quantisation, verify success rate before trusting it |
Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.
| Inference memory, reported | 32.54 GB source |
The 2B model card lists 32.54 GB of GPU memory for 720p, 16 FPS video to world generation and single generation times of 228.8 s on an H100 SXM, 123.9 s on a B200 and 2,567.1 s on an L40S. Only BF16 on Ampere, Hopper and Blackwell GPUs is tested. The 14B card publishes no memory or timing figures yet.
Reported results
Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.
Simulation
No results in this category are published for this model.
Real world
No results in this category are published for this model.
Fine tuned tasks
No results in this category are published for this model.
On real hardware
No real robot rollouts are reported for the world model itself. The robotics evidence is offline: on 100 Bridge test episodes the 2B action-conditioned model reaches PSNR 24.95, SSIM 0.85 and FVD 146, against 21.14, 0.82 and 190 for Cosmos-Predict1-7B, and the 14B GR00T Dreams variant scores 91.8 on object and 69.0 on environment generalization in DreamGen Bench with GPT-4o as judge.
Fine tuning it yourself
Robot use goes through post-trained variants rather than the base model: action-conditioned, multiview and policy checkpoints of the 2B model are published, and the GR00T Dreams workflow post-trains Cosmos world models on robot video. Code is Apache-2.0 in the cosmos-predict2.5 repository.
Where it helps, where it does not
Strengths
- Open weights at 2B and 14B under the NVIDIA Open Model License, which allows commercial use according to the 14B model card.
- Robot variants are published as checkpoints: action-conditioned, multiview and a policy variant post-trained on LIBERO and RoboCasa.
- Clear step over its predecessor on action-conditioned robot video: FVD 146 against 190 for Cosmos-Predict1-7B on Bridge, with a 2B instead of a 7B model.
- One model covers text, image and video conditioning, so the same weights serve data generation and policy evaluation.
Limits
- Slow and memory heavy: 228.8 s per 720p clip on an H100 SXM and 32.54 GB of GPU memory for the 2B model.
- No real robot result for the world model itself; the robotics numbers are video metrics and judged benchmarks.
- The base model does not output actions; robot use needs a post-trained variant or an inverse dynamics model.
- The 2B weights require accepting the licence on Hugging Face.
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://arxiv.org/abs/2511.00062
- https://arxiv.org/html/2511.00062v2
- https://github.com/nvidia-cosmos/cosmos-predict2.5
- https://huggingface.co/nvidia/Cosmos-Predict2.5-2B
- https://huggingface.co/nvidia/Cosmos-Predict2.5-14B
Entry last checked 2026-11-12.
Try a policy on a real arm
A real SO-100 runs on /live and can be driven from the browser when it is online. Five of the models in this arena can be fine tuned on a dataset you record with your own arm.