Meta has released V-JEPA 2, a 1.2 billion-parameter world model trained on video for understanding and prediction in the physical world and zero-shot robot planning, Meta said in a blog post on 11 June 2025.
V-JEPA 2 is built on Meta's Joint Embedding Predictive Architecture (JEPA). Pre-training used more than 1 million hours of video and 1 million images, followed by action-conditioned training.
Meta said training with only 62 hours of robot data produced a model that can be used for planning and control. Using visual subgoals, V-JEPA 2 achieved success rates of 65% to 80% for picking and placing new objects in unseen environments.
Meta also released three benchmarks for physical reasoning from video: IntPhys 2, Minimal Video Pairs (MVPBench) and CausalVQA. It said humans score 85% to 95% on them, well ahead of top models including V-JEPA 2.
The model artifacts are available on GitHub and Hugging Face.