Meta's V-JEPA 2 trains a world model on video to guide robot actions

What Meta actually shipped

On June 11, 2025, Meta introduced V-JEPA 2, a world model trained on video that the company positions as a successor to the original V-JEPA released last year. The stated goal is physical reasoning: giving AI agents an internal model of how the world will respond to their actions.

Meta frames three capabilities as the core of a world model — understanding, predicting, and planning. The first V-JEPA covered understanding and prediction from video. V-JEPA 2 is presented as improving both to the point where robots can act on the model's predictions.

Alongside the model, Meta released three new benchmarks meant to let other researchers measure how well their own models reason about the physical world from video. That pairing — a model plus the yardsticks to evaluate models in the same domain — is the more consequential part of the release.

The robot demonstration is narrow, and Meta says so

The concrete evidence Meta offers is a lab result: robots using V-JEPA 2 to reach for objects, pick them up, and place them in a new location. These are the canonical primitives of robotic manipulation, and the announcement describes them plainly rather than as finished products.

The phrase Meta uses is that robots can 'interact with unfamiliar objects and environments to complete a task.' That framing matters. Generalizing to unfamiliar objects is the hard part of manipulation; a model that learned interaction patterns from video rather than from task-specific teleoperation data is a different bet than most robotics pipelines make.

But the source stops at reach, pick, and place in Meta's own labs. There is no claim about success rates, task complexity, or deployment outside controlled conditions. Read literally, this is an early capability demonstration attached to a model release.

Learning physics from video instead of from labeled action data

We trained V-JEPA 2 using video, which helped the model learn important patterns in the physical world, including how people interact with objects, how objects move in the physical world and how objects interact with other objects.Montana Labs

The bet behind this approach is that passive video is abundant and physics is consistent, so a model can absorb how objects move and collide without being explicitly labeled or driven through every action. The tennis-ball and hockey-puck analogies in the announcement are Meta's way of describing prediction of future state rather than recognition of current state.

This is a distinct lineage from language-model-driven agents that reason in text. Here the reasoning substrate is a learned representation of visual dynamics, and the 'thinking before acting' in the title refers specifically to predicting outcomes before executing a physical action.

The benchmarks are the part worth watching for applied teams

For anyone building on physical-world AI, the three benchmarks are more actionable than the model itself. A model release is a single artifact; shared evaluation criteria shape how an entire field measures progress, and whoever defines the benchmark tends to define what counts as success.

Meta explicitly ties this to accelerating outside research, offering both 'the best models and benchmarks' to the community. Teams evaluating world models for robotics should treat these benchmarks as a claim about what physical reasoning means — one to adopt or contest, not accept unexamined.

The specific implication: V-JEPA 2 signals that Meta is competing on the video-to-action world model, and by publishing the yardsticks alongside the model it is trying to set the terms on which everyone else's physical-reasoning models get judged. That standard-setting move outlasts the reach-pick-place demo it arrived with.