V-JEPA 2: Meta's Video World Model That Understands, Predicts — and Plans
— by Vivax
Meta AI released V-JEPA 2, a 1.2-billion-parameter world model trained on over a million hours of video that sets state-of-the-art marks in visual…
On June 11, 2025, Meta AI released V-JEPA 2 — a 1.2-billion-parameter world model trained primarily on video, presented as a step toward advanced machine intelligence (AMI): AI systems that learn about the world the way humans do, predict how it will evolve, and plan how to act in it. The framing is Yann LeCun's world-model thesis made concrete: before acting, an agent should be able to understand what it observes, predict what happens next — including the consequences of its own actions — and plan a sequence of actions toward a goal. The release is deliberately open, with code and checkpoints available for both commercial and research use.
Architecturally, V-JEPA 2 stays faithful to the joint-embedding predictive architecture: an encoder turns raw video into embeddings that capture the semantic state of the world, and a predictor forecasts future embeddings — predictions live in latent space, not pixels, so the model learns dynamics without wasting capacity on rendering every leaf and shadow. Training happens in two stages. The first is actionless self-supervised pre-training on more than one million hours of video and a million images. That alone carries remarkably far: with lightweight readouts on frozen features, V-JEPA 2 posts top marks on Something-Something v2 motion understanding, a new state of the art on Epic-Kitchens-100 action anticipation — predicting what someone will do one second from now — and, aligned with a language model, state-of-the-art results on video question-answering benchmarks like Perception Test and TempCompass.
The second stage gives the model a body. Feeding just 62 hours of robot interaction data from the open DROID dataset into the same predictive objective — this time conditioned on actions — turns V-JEPA 2 into a planner. Deployed zero-shot on robot arms in Meta's labs, with no training data from that robot or that room, it plans by imagination: given a goal image, the robot uses the predictor to simulate the consequences of candidate actions, rates them by how close they land to the goal, executes the best one, and re-plans at every step. With visual subgoals for longer tasks, it achieves 65–80% success picking and placing objects it has never seen, in environments it has never entered.
Meta paired the model with a reality check: three open benchmarks probing whether video models genuinely understand the physical world. IntPhys 2 uses the violation-of-expectations paradigm from developmental psychology — paired videos identical until one breaks physics — and finds humans near-perfect while current video models sit at or close to chance. MVPBench asks multiple-choice questions over minimal-change video pairs, so a model only scores when it answers both twins correctly, foiling shortcut learning. CausalVQA probes physical cause-and-effect: counterfactuals, anticipation, planning — where models that fluently describe what happened stumble on what would have happened. Across all three, humans score 85–95%; the gap to today's best models, V-JEPA 2 included, is published openly on a Hugging Face leaderboard.
V-JEPA 2 is the clearest public demonstration of the thesis Vivax is built on. Our clinical world model belongs to the same JEPA lineage — learn predictive representations in latent space, then read out understanding, prediction, and planning from one backbone. The understand-predict-plan loop is the clinical loop: recognize the patient's state, forecast its trajectory, plan the intervention. And the 62-hour result carries a lesson we care about deeply — a world model pre-trained on abundant passive data needs astonishingly little interaction data to become actionable, which in medicine means learning the dynamics from the health data that already exists, then specializing to each clinical task with the scarce labeled data hospitals can actually provide. The benchmark gap is the honest part: physical intuition is not solved — which is exactly why world models must stay grounded in measured reality. That is where we build.