Introducing the V-JEPA 2 world model and new benchmarks for physical reasoning

TL;DR

Meta’s June-2025 announcement of V-JEPA 2 — a 1.2B-parameter video world model trained with self-supervised mask-denoising under the Joint Embedding Predictive Architecture (JEPA), released with code, checkpoints, and three new physical-reasoning benchmarks (IntPhys 2, MVPBench, CausalVQA). Training is two-stage: actionless pre-training on >1M hours of video and 1M images, followed by action-conditioned post-training on just 62 hours of robot data (DROID). The result is a frozen-encoder world model that supports zero-shot model-predictive control on real robots for pick-and-place (65–80% success on unseen objects in unseen environments), plus SOTA on Something-Something v2 action recognition and Epic-Kitchens-100 action anticipation via attentive probes. This is the foundational V-JEPA 2 release — predecessor to V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning and the version used as a frozen reward model in Inference-time Physics Alignment of Video Generative Models with Latent World Models.

Key claims

V-JEPA 2 is a 1.2B-parameter video world model built on the JEPA self-supervised framework (encoder + predictor in latent space, no pixel reconstruction) [post body].
Pre-training uses >1M hours of video and 1M images from diverse sources with no human annotation; an attentive read-out on top of frozen encoder features achieves SOTA on Something-Something v2 (motion understanding) and a read-out over frozen encoder + predictor features sets SOTA on Epic-Kitchens-100 action anticipation [post body].
Aligning V-JEPA 2 with a language model yields SOTA on video QA benchmarks Perception Test and TempCompass [post body].
Action-conditioned post-training requires only 62 hours of robot data; the second-stage predictor learns to condition on control actions and is then usable for model-predictive control [post body].
Zero-shot deployment: trained on the open-source DROID dataset, deployed directly on lab robots with no robot-specific or environment-specific fine-tuning; pick-and-place success rates of 65–80% on new objects in new environments using image-goal MPC, with visual subgoals for longer-horizon tasks [post body].
IntPhys 2 — a violation-of-expectation benchmark built with a game engine: pairs of identical-prefix videos that diverge on a physics-breaking event; humans near-perfect, current video models at or near chance [post body].
MVPBench — multiple-choice video-language QA with minimal-change pairs (same question, visually similar video, opposing answer); a model only scores credit if it gets both members of a pair correct, mitigating textual/visual shortcut exploitation [post body].
CausalVQA — counterfactual / anticipation / planning questions on physical-world videos; models are increasingly capable on “what happened” but still lag on “what could have happened” and “what might happen next” [post body].
Aggregate result on all three benchmarks: humans 85–95% accuracy vs. top models including V-JEPA 2 at substantially lower — a quantified human-machine gap on intuitive physical reasoning [post body].
Stated next directions: hierarchical JEPA models that learn and plan across multiple temporal/spatial scales, and multimodal JEPA models extending to audio and touch [post body].

Method

V-JEPA 2 stacks two training stages on the JEPA scaffold. Stage 1 (actionless pre-training) trains an encoder and a latent predictor by masking parts of a clip and predicting the latent representations of the masked region from visible context — no pixel reconstruction, no language supervision. The corpus is >1M hours of video plus 1M images. After this stage, the model can already predict how the world evolves but without conditioning on agent actions. Stage 2 (action-conditioned training) feeds the predictor an additional action stream alongside visual context, using the open-source DROID robot dataset (just 62 hours used) to learn an action-conditional dynamics model in latent space. At deployment, the encoder embeds the current observation and an image-goal, and the predictor rolls out candidate action sequences in latent space; an MPC loop picks the action whose predicted latent is closest to the goal latent, re-plans every step, and chains visual subgoals for longer horizons.

Results

Pre-trained encoder + attentive read-out: SOTA on Something-Something v2 (motion understanding). Pre-trained encoder + predictor + attentive read-out: SOTA on Epic-Kitchens-100 1-second action anticipation. Language-aligned: SOTA on Perception Test and TempCompass video QA. Robot deployment on Franka arms in Meta labs, zero-shot from DROID-only training: 65–80% pick-and-place success on previously unseen objects in previously unseen environments using image-goal MPC and visual subgoaling. Benchmark gap: humans hit 85–95% on IntPhys 2 / MVPBench / CausalVQA while top video models (V-JEPA 2 included) sit much lower — IntPhys 2 specifically at or near chance for current video models. No exact V-JEPA 2 scores are quoted in the blog (full numbers are in the accompanying technical report).

Why it’s interesting

This is the canonical V-JEPA 2 launch that the wiki has been citing indirectly — the foundation under V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning (which extends the masked-prediction loss to all tokens to fix dense-feature quality) and the model used frozen as a physics-plausibility reward in Inference-time Physics Alignment of Video Generative Models with Latent World Models. Two design choices stand out for Luma. First, the 62-hour action-conditioned post-training: a frozen pre-trained latent world model can absorb a tiny robot-specific dataset and still drive zero-shot MPC on new objects/environments — a much cheaper recipe than VLA approaches that demand robot-instance-specific data. Second, the explicit decoupling of world-model representation from policy: the same JEPA encoder serves SSv2 action recognition, Epic-Kitchens anticipation, language-grounded VQA, and MPC, contrasting with the World Foundation Models cluster’s other axis where Cosmos-style stacks emphasize generative rollout. Combined with the IntPhys 2 / MVPBench / CausalVQA release, this blog is also the first published artifact in the wiki that explicitly couples a world-model release to a probing benchmark suite designed to expose physical-reasoning failures — a methodological pattern other WFM releases (Genie 3, LingBot-World) so far don’t follow.