V-JEPA 2

Released June 11, 2025, V-JEPA 2 is Meta's second-generation video joint-embedding predictive architecture: a 1.2-billion-parameter vision encoder trained on more than a million hours of video plus a million images, learning to predict the latent representation of masked video regions rather than pixels. The action-conditioned variant, V-JEPA 2-AC, adds a ~300M-parameter predictor that forecasts future latent states given candidate actions, enabling model-predictive control of robot arms with only about 62 hours of real robot interaction data.

Meta released the code and model checkpoints publicly, and V-JEPA 2 set new marks on physical-reasoning benchmarks (IntPhys 2, CausalVQA). The honest limitation is that latent-space planning is still far from general autonomy: reported zero-shot manipulation works in constrained lab settings, and the latent predictions are not human-interpretable, which complicates debugging and safety analysis for production robotics.