nvidia-cosmos/cosmos-predict2.5

Health v2 · maintenance only

Adoption not part of the score

1372 stars · 194 forks observed · 2026-09-23

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

NVIDIA Cosmos-Predict2.5 is a family of world foundation models (WFMs) that generate video predictions of future world states for physical AI applications like autonomous vehicles and robotics. The repository provides inference and post-training code, though it is now superseded by Cosmos 3 and receives only limited maintenance.

Use cases

• generate future-state video predictions for robot training
• post-train a world model on custom robotics data
• synthesize training video for autonomous vehicle simulation
• action-conditioned video generation for embodied agents
• distill a large video world model for faster inference
• build physical AI data pipelines with synthetic video

When to choose

• you need video-based world prediction models for robotics or AV research
• you want to post-train or distill Cosmos Predict2.5 checkpoints
• you need reproducible recipes tied to the Predict2.5 model family

When to avoid

• you want the latest capabilities - migrate to NVIDIA Cosmos 3
• you need a lightweight video generation tool without GPU infrastructure
• you expect active development or new features in this repo

Facets

library · maturity maintenance

machine-learning deep-learning video-processing llm-training simulation artificial-intelligence machine-learning deep-learning robotics simulation python world-foundation-models video-generation physical-ai post-training diffusion-models autonomous-vehicles robotics-simulation nvidia video linux gpu docker

Member repositories

For agents

markdown · JSON · MCP: product_card(name="nvidia-cosmos/cosmos-predict2.5")