V-JEPA 2

Architecture

Self-supervised joint-embedding predictive video world model (~1.2B params); action-conditioned V-JEPA 2-AC variant for robot control.

Key facts

Model

V-JEPA 2: self-supervised video WORLD MODEL, ~1.2B params, trained on >1M hours of video + 1M images (released Jun 11 2025). OPEN: code + checkpoints on GitHub/Hugging Face; arXiv 2506.09985. Plus new physical-reasoning benchmarks.

Robotics variant

V-JEPA 2-AC: action-conditioned variant post-trained on ~62 hrs of unlabeled robot video (open-source DROID dataset) for zero-shot control.

Deployed wiring: NONE (research only)

Demonstrated zero-shot pick-and-place of new objects in new environments at 65-80% success ON ROBOTS IN META'S LABS. No commercial robotics product, no named customers, no robot-vendor wiring. The cleanest research/world-model case. No registry robot should be ModelBrain-wired to it (verifiable-only wiring; no real-robot product exists).

Cap-flag

Demonstrated capability in Meta labs, NOT deployed at customers. Whether 65-80% lab numbers generalize beyond the tested arms is unverified; it is short-horizon reach/grasp/place, not long-horizon manipulation.

Developed by (1)

• Meta

Common questions

What is V-JEPA 2?

Meta's Fundamental AI Research lab released V-JEPA 2 on June 11, 2025, a self-supervised video world model of about 1.2 billion parameters trained on more than a million hours of video and a million images, in Yann LeCun's Joint-Embedding Predictive Architecture line. It is open, with code and checkpoints on GitHub and Hugging Face and an arXiv paper, and shipped alongside new physical-reasoning benchmarks. An action-conditioned variant, V-JEPA 2-AC, is post-trained on roughly 62 hours of unlabeled robot video from the open-source DROID dataset to enable zero-shot control. The registry records it at research maturity, the cleanest research world-model case in this expansion: it was demonstrated doing zero-shot pick-and-place of new objects in new environments at 65 to 80 percent success on robots in Meta's own labs, but there is no commercial robotics product, no named customers, and no robot-vendor wiring, so no registry robot should be wired to it under the verifiable-only-wiring rule since no real-robot product exists. Its strong open-artifact case of published weights, paper, and benchmarks clears the registrable bar easily, while the deployed-wiring caveat is explicit: this is demonstrated lab capability, not deployment, the lab success numbers are not shown to generalize beyond the tested arms, and the manipulation is short-horizon rather than long-horizon.

What type of AI is V-JEPA 2?

V-JEPA 2 is a world model on the DEPLOY registry. It is open-source / open-weights.

Who developed V-JEPA 2?

V-JEPA 2 is credited to Meta on the DEPLOY registry. Each developer attribution is verified via primary sources.

Which robots run on V-JEPA 2?

No robot models on the DEPLOY registry are recorded as running V-JEPA 2. DEPLOY wires brain-to-model connections only when the wiring is verifiable from primary sources; absence may reflect pre-deployment or unverified manufacturer claims.

Is V-JEPA 2 open source?

Yes. V-JEPA 2 is recorded as open-source / open-weights on the DEPLOY registry, meaning model weights or source code are publicly available.

What is V-JEPA 2's maturity stage?

V-JEPA 2 is at the research stage on the DEPLOY maturity ladder. Research stage means active development without commercial deployments on file.

Sources (3)

1. https://ai.meta.com/blog/v-jepa-2-world-model-benchmarks/ ↗
2. https://arxiv.org/abs/2506.09985 ↗
3. https://venturebeat.com/ai/metas-new-world-model-lets-robots-manipulate-objects-in-environments-theyve-never-encountered-before ↗