NVIDIA Cosmos

Robotics and autonomous-vehicle teams use Cosmos as a source of physics-aware synthetic video and embodied reasoning. The platform spans three open model families plus tooling: Cosmos-Predict, now at version 2.5, simulates the future state of a scene as video, Cosmos-Transfer 2.5 renders high-quality world simulations conditioned on spatial control inputs such as depth or segmentation maps, and Cosmos-Reason 2 handles physical common sense, producing embodied decisions in natural language. Around the models sit Cosmos-RL, a reinforcement learning framework built for physical AI, and the Cosmos-Cookbook of post-training scripts for adapting the checkpoints to a specific robot or driving domain. Code across the primary repositories is Apache 2.0, with model weights distributed under the permissive NVIDIA Open Model License, so the whole stack self-hosts on your own GPUs; earlier Predict1, Transfer1, and Reason1 generations remain available alongside the current releases. NVIDIA positions the family as trained infrastructure for physical AI that developers post-train on their own data rather than a consumer video toy.