V-JEPA 2
V-JEPA 2 (Video Joint Embedding Predictive Architecture 2) is an open-source video world model released by Meta AI on June 11, 2025 that learns to understand, predict, and plan in the physical world by watching internet video. It is a 1.2 billion parameter model pretrained via self-supervised learning on more than one million hours of internet video together with roughly one million images, and it is the most prominent published embodiment of Yann LeCun's Joint Embedding Predictive Architecture (JEPA) paradigm for autonomous machine intelligence. [1][2] Unlike generative world models that synthesize pixels, V-JEPA 2 predicts future states in a learned representation space, which Meta says makes it 30 times faster than NVIDIA's Cosmos model. [9] A companion variant, V-JEPA 2-AC, is post-trained on under 62 hours of unlabeled teleoperated robot footage from the public Droid dataset and enables zero-shot robot control on Franka Panda arms in laboratories the model has never seen. [1][2]
The model was published in the arXiv preprint V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning (arXiv:2506.09985) by a team led by Mido Assran, including Yann LeCun, Michael Rabbat, Nicolas Ballas, Adrien Bardes, and David Fan. [1] Meta released code, multiple checkpoints, and evaluation probes under an MIT license through the facebookresearch/vjepa2 GitHub repository and a companion Hugging Face collection. [4] Meta describes the work as "meaningful progress toward our ultimate goal of developing advanced machine intelligence (AMI)," arguing that world models are "essential to building AI agents that can think before they act." [3]
V-JEPA 2 is notable for three reasons. First, it predicts future states in a learned latent representation space rather than in pixel space, in keeping with LeCun's long argued thesis that generative pixel prediction wastes capacity on perceptually irrelevant detail. [6] Second, it produces a single visual backbone that achieves competitive or state-of-the-art results across motion understanding, action anticipation, and video question answering when paired with a frozen probe or a language model. [1] Third, by demonstrating zero-shot pick-and-place on a real robot trained only on internet video plus a small amount of unlabeled teleoperation data, it is presented by Meta as evidence that a non-generative world model can support real robot planning without environment-specific data collection. [1][2]
What is the JEPA paradigm behind V-JEPA 2?
The JEPA framework was introduced by Yann LeCun in his 2022 position paper A Path Towards Autonomous Machine Intelligence. [6] In that paper LeCun proposed that systems aiming for human-like reasoning should learn predictive world models in an abstract embedding space rather than by reconstructing raw inputs. A JEPA model contains an encoder that maps an input into a representation, a predictor that maps from one such representation to another conditioned on context or actions, and a training objective that aligns the predicted representation with a target representation produced by a target encoder. Because the prediction is in latent space, the system is free to discard unpredictable low-level detail such as exact pixel values, background noise, and viewing angle, and instead capture the structural regularities that govern how a scene evolves.
Meta's research group released several JEPA implementations before V-JEPA 2. I-JEPA, presented in 2023, applied the framework to still images and showed that masked-region prediction in latent space could produce strong visual representations without contrastive learning, data augmentation, or pixel reconstruction. [7] V-JEPA, released in early 2024, extended the same recipe to short video clips, training a vision transformer to predict the embeddings of masked spatiotemporal tubes given the embeddings of the visible portion. [8] V-JEPA 2 scales this approach in both data and model size (from about 2 million to 22 million videos, and to a 1 billion parameter ViT-g encoder) [1], adds a second post-training stage focused on action-conditioned prediction, and introduces three new benchmarks aimed at separating genuine physical understanding from textual or statistical shortcuts. [2]
JEPA lineage
| Model | Year | Modality | Pretraining data | Parameter count | Primary contribution |
|---|---|---|---|---|---|
| I-JEPA | 2023 | Still images | ImageNet-1K (1.3M images) and ImageNet-22K | ViT-H/14 (632M) and ViT-G/16 | First JEPA-based vision model; latent-space masked prediction |
| V-JEPA | 2024 | Short videos | ~2M public videos | ViT-L and ViT-H | Extends latent masked prediction to spatiotemporal tubes |
| V-JEPA 2 | June 2025 | Long-form internet video plus images | >1M hours of video plus ~1M images | ViT-g encoder over 1B params; 1.2B world-model configuration | Internet-scale self-supervised video pretraining for understanding, prediction, and planning |
| V-JEPA 2-AC | June 2025 | Robot teleoperation | <62 hours of unlabeled Droid footage | 300M action-conditioned predictor over frozen ViT-g encoder | Latent action-conditioned world model enabling zero-shot robot planning |
How is V-JEPA 2 built?
V-JEPA 2 follows the canonical JEPA template with two transformer modules. As Meta describes it, the architecture has "two main components: an encoder, which takes in raw video and outputs embeddings that capture useful semantic information about the state of the observed world, and a predictor, which takes in a video embedding and additional context about what to predict and outputs predicted embeddings." [2] The encoder is a vision transformer applied to a video clip that has been tokenized into spatiotemporal patches; the predictor is a separate transformer that takes the encoder's output for a context region and a set of position queries and outputs predicted embeddings for the masked region. The training objective is a regression loss between the predicted embeddings and target embeddings produced by an exponential-moving-average copy of the encoder. To prevent representational collapse, V-JEPA 2 inherits the design choices that have proved stable across the JEPA family, including stop-gradient through the target encoder.
The released V-JEPA 2 encoder uses a ViT-g backbone with over one billion parameters, and the full world-model configuration combining the encoder with its predictor is described by Meta as a 1.2 billion parameter model. [2] The video-and-image pretraining corpus, named VideoMix22M, comprises roughly 22 million samples drawn from public sources. [1] Most pretraining uses 16-frame clips at 256 by 256 resolution, followed by a shorter cooldown phase on clips of up to 64 frames at 384 by 384 resolution. [1]
For V-JEPA 2-AC, the action-conditioned variant, Meta keeps the V-JEPA 2 encoder frozen and trains a new 300 million parameter transformer predictor with block-causal attention. [1] This predictor takes a sequence of past encoded frames plus a sequence of low-dimensional action vectors (end-effector control actions, together with end-effector states) and autoregressively predicts the encoder's representation of the next frame. Because both the past states and the predicted states live in the same latent space, planning at test time can be performed entirely in that space without ever generating pixels. Image-goal planning is implemented using model-predictive control: candidate action sequences are rolled out through the action-conditioned predictor, the resulting predicted latent is compared with the encoded goal image, and the cross-entropy method is used to refine the action distribution. [1]
How was V-JEPA 2 trained?
V-JEPA 2 pretraining proceeds in a single self-supervised stage on the combined video-and-image corpus. The paper describes pretraining on "a video and image dataset comprising over 1 million hours of internet video," built from four video sources (Something-Something v2, Kinetics, HowTo100M and YT-Temporal-1B) plus ImageNet images, each sampled with its own weight. [1][10] During pretraining the model is given a tokenized video clip from which a substantial fraction of patches are masked, and it is trained to predict the embedding of those masked tubes. Crucially, the loss is computed in feature space against the target encoder rather than against the raw pixel values, which the team has repeatedly argued is essential for learning representations that are robust to nuisance variation. [6]
The action-conditioned post-training stage for V-JEPA 2-AC uses less than 62 hours of unlabeled video drawn from the Droid dataset, the open-source teleoperation corpus collected from a fleet of Franka Panda arms across many institutions. [1] "Unlabeled" here means that the post-training does not depend on natural language task descriptions or reward labels, only on the synchronized pairing of robot proprioception and video. The total compute and wall-clock cost of the action-conditioned post-training is small compared to internet-scale pretraining, reflecting the JEPA philosophy that most knowledge about the physical world should already be present in the encoder after large-scale passive observation.
For downstream evaluations the team uses two main protocols. The first freezes the encoder and trains a lightweight attentive read-out probe for classification or anticipation tasks. The second aligns the encoder with a large language model through a projector and a multi-stage visual instruction-tuning process, with the encoder either frozen or fine-tuned, to support video question answering. [1]
How well does V-JEPA 2 perform on benchmarks?
V-JEPA 2 was evaluated across a wide suite of established video understanding benchmarks. Two of the headline results sit on Something-Something v2, which tests fine-grained motion understanding, and Epic-Kitchens-100, which tests action anticipation one second into the future. The paper reports that V-JEPA 2 "achieves strong performance on motion understanding (77.3 top-1 accuracy on Something-Something v2) and state-of-the-art performance on human action anticipation (39.7 recall-at-5 on Epic-Kitchens-100) surpassing previous task-specific models." [1] Aligned with a large language model, it also sets state-of-the-art video question-answering numbers at the 8 billion parameter scale. [1]
Headline V-JEPA 2 benchmark scores
| Benchmark | Capability tested | V-JEPA 2 score | Notes |
|---|---|---|---|
| Something-Something v2 | Fine-grained motion classification | 77.3 top-1 accuracy | Frozen encoder with attentive read-out |
| Epic-Kitchens-100 | Action anticipation at one second horizon | 39.7 recall at 5 | New state of the art; surpasses previous task-specific models |
| Perception Test | Multimodal video question answering | 84.0 | ViT-g encoder aligned with Llama 3.1 8B (8B scale); test set, with SFT |
| TempCompass | Temporal reasoning in video QA | 76.9 | ViT-g encoder aligned with Llama 3.1 8B (8B scale) |
Meta also introduced three new benchmarks alongside V-JEPA 2 to probe physical understanding more rigorously, describing them as "three new benchmarks to help the research community evaluate how well their existing models learn and reason about the world using video." [2] IntPhys 2 evaluates the ability of a model to detect physically implausible events, a task on which humans achieve near-perfect accuracy while current video models score at or close to chance; across all three new benchmarks, Meta reports human accuracy of 85 to 95 percent. [2] MVPBench, or Minimal Video Pairs Bench, presents the same video question with two nearly identical clips that differ in the answer, removing many of the surface-level shortcuts that video question-answering models can otherwise exploit. [2] CausalVQA targets cause-and-effect reasoning, counterfactual questions, and short-horizon anticipation. [2] On these new benchmarks V-JEPA 2 and other top foundation models still trail human performance by a wide margin, which the authors argue is exactly the headroom that motivates further research on world modeling.
How well does V-JEPA 2-AC control real robots?
The action-conditioned variant was deployed zero-shot on Franka Panda arms in two different laboratories, neither of which contributed data to the encoder or predictor training. [1] Tasks were specified as goal images (a single goal image for short-horizon tasks, a series of visual subgoals for pick-and-place); the controller used model-predictive control through the latent predictor to pick action sequences, computing each action in roughly 16 seconds. [1] Averaged across the two labs, reported success rates were 100 percent on reach tasks, 65 percent (cup) and 25 percent (box) on grasping, and 80 percent (cup) and 65 percent (box) on pick-and-place. [1][2] The results are notable because the policy is goal-conditioned rather than language-conditioned, has never seen the specific robots or objects in question, and never receives a reward signal. The paper underscores that this "is achieved without collecting any data from the robots in these environments, and without any task-specific training or reward." [1]
How does V-JEPA 2 differ from other world models?
V-JEPA 2 arrived in the middle of a broader shift among large research labs toward what are loosely called world foundation models. These systems share the goal of letting agents predict and plan in a learned model of the physical world, but they differ sharply in what they actually predict and how they are intended to be used. The table below summarizes how V-JEPA 2 compares to two prominent contemporaries, Google DeepMind's Genie line and NVIDIA's Cosmos suite.
Comparison with other 2025 world models
| System | Developer | Architecture style | What it predicts | Primary use case | Approximate scale | Openness |
|---|---|---|---|---|---|---|
| V-JEPA 2 | Meta AI | Joint embedding predictive (non-generative) | Future latent representations of video | Understanding, prediction, robot planning | 1.2B parameters; >1M hours video pretraining | Open weights and code under MIT |
| Genie 3 | Google DeepMind | Autoregressive generative simulator | Pixel-level next frames conditioned on actions | Interactive, navigable simulated worlds | Not disclosed | Closed; research preview at launch, then available to US Google AI Ultra subscribers through Project Genie from January 2026 [12] |
| NVIDIA Cosmos | NVIDIA | Generative world foundation model family (Predict, Transfer, Reason) | Pixel-level video and physics-aware synthetic data | Simulation, synthetic data for robotics and AV | 7B and 14B parameter variants | Open weights with permissive license |
Meta has said that V-JEPA 2 is 30 times faster than NVIDIA Cosmos, though TechCrunch noted Meta may be using different benchmarks than NVIDIA [9]; in the paper's robot-planning comparison, V-JEPA 2-AC needed 16 seconds per action against 4 minutes for a fine-tuned Cosmos model. [1] The gap is consistent with the architectural difference: V-JEPA 2 predicts a small set of latent features per frame, whereas pixel-generative systems must produce a full image. [2] The trade-off is that V-JEPA 2 cannot be used to render a video for a human viewer; its outputs are only meaningful when fed back into the planner or into a downstream probe. World Labs, founded in 2024 to build 3D world models, occupies yet another point in the design space, focusing on generating navigable 3D scenes from images rather than predicting video futures.
The broader debate that V-JEPA 2 is intended to advance, and that LeCun has championed publicly for years, is whether the next generation of foundation models should be generative or predictive in latent space.
Is V-JEPA 2 open source?
Meta released V-JEPA 2 through the facebookresearch/vjepa2 repository on GitHub. [4] The majority of the code is licensed under the MIT License, with a small number of utility files under Apache 2.0. Pretrained encoder weights for all model sizes are available, along with the action-conditioned predictor checkpoint trained from the ViT-g encoder. The release also includes evaluation probes for Something-Something v2, Diving48, and Epic-Kitchens-100, plus integration with PyTorch Hub and a companion Hugging Face collection. [4]
In March 2026 Meta released V-JEPA 2.1, which changes the training recipe to learn dense, temporally consistent features: its dense predictive loss counts both visible and masked tokens, and it adds deep self-supervision across intermediate encoder layers and multi-modal tokenizers for images and video. The lineup comprises ViT-g/16 (1 billion parameters) and ViT-G/16 (2 billion) encoders plus distilled ViT-B/16 (80 million) and ViT-L/16 (300 million) models, all at 384 resolution, and the paper reports a 20-point gain in real-robot grasping success over V-JEPA 2-AC. [4][13]
Meta also published a public Hugging Face leaderboard for the newly introduced IntPhys 2, MVPBench, and CausalVQA benchmarks. The intent, as expressed in the original blog post, is to encourage the community to develop and openly evaluate world models on tasks that go beyond conventional video classification. [2]
Why does V-JEPA 2 matter for embodied AI?
For the embodied AI community, the most significant aspect of V-JEPA 2 is the V-JEPA 2-AC demonstration. Prior work on learned robot policies typically required tens of thousands of robot trajectories collected on the specific embodiment and in similar environments, or relied on heavily engineered simulation pipelines. V-JEPA 2-AC reuses representations learned from internet video, post-trains for a short time on a publicly available teleoperation corpus, and is then deployed on robots in unfamiliar laboratories without any additional data collection. [1] This pattern, sometimes referred to as zero-shot embodied transfer, is widely seen as a prerequisite for scaling general-purpose robot intelligence in the same way that internet-scale pretraining scaled language models.
The results also reframe the role of simulation in robotics. Generative world models such as Genie and Cosmos are positioned as ways to synthesize training data for downstream controllers, effectively replacing or augmenting classical simulators. V-JEPA 2 instead skips the generation step and uses the learned predictor directly inside the control loop. Both approaches are likely to coexist, and several research groups have begun exploring hybrid systems that use pixel-generative world models for data augmentation and JEPA-style latent predictors for closed-loop control.
Open questions remain. In the June 2025 paper, V-JEPA 2-AC is evaluated on tabletop manipulation with a relatively short planning horizon and a single arm; it has not yet been demonstrated on dexterous bimanual tasks, mobile manipulation, or long-horizon plans that require explicit task decomposition. The benchmarks introduced alongside the model also reveal substantial gaps between V-JEPA 2 and human performance on physical plausibility detection and causal reasoning, suggesting that the JEPA recipe will need to scale further or be combined with more structured objectives before it can match human-level physical intuition. [2]
Reception
Trade coverage focused on the robotics angle and Meta's performance claims. TechCrunch reported Meta's claim that V-JEPA 2 is 30 times faster than NVIDIA's Cosmos, noting that Meta may be evaluating its models on different benchmarks than NVIDIA. [9] MarkTechPost described it as "a scalable open-source world model" and highlighted the 16-second planning time against 4 minutes for Cosmos. [10] The Robot Report placed it alongside Google DeepMind's Genie and World Labs as part of a wider push by technology companies into world models. [11]
Continue reading
- V-JEPAV-JEPA (Video Joint Embedding Predictive Architecture) is a self-supervised video model from Meta AI that learns by predicting masked regions of a video in an abstract latent representation space rather than…Mentioned 29 times above
- World modelA world model is an artificial intelligence system that learns an internal representation of how an environment works, enabling it to predict future states, simulate the consequences of actions, and support…Mentioned 11 times above
- Yann LeCunYann Andre LeCun (born July 8, 1960) is a French-American computer scientist, a pioneer of convolutional neural networks (CNNs) and deep learning, and a 2018 ACM A.M. Turing Award laureate often called one of…Mentioned 3 times above
- Joint Embedding Predictive ArchitectureJoint Embedding Predictive Architecture (JEPA) is a family of self-supervised, non-generative neural network architectures proposed by Yann LeCun in his June 2022 position paper A Path Towards Autonomous…Mentioned twice above
- I-JEPAI-JEPA (Image-based Joint-Embedding Predictive Architecture) is a self-supervised learning method for computer vision developed by Meta AI.Mentioned twice above
- Embodied AIEmbodied AI is an area of artificial intelligence concerned with agents whose perception, action, learning, and reasoning are coupled through interaction with an environment.Mentioned twice above
More from See also
- Meta AI
- Self-Supervised Learning
- Vision Transformer
- Model Predictive Control
- Genie 3
- NVIDIA Cosmos
- World Labs
- Rerun (rerun.io)
Topics:AI ModelsComputer VisionOpen Source AIRoboticsWorld Models
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21Assran, M. et al. (2025). *V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning*. arXiv:2506.09985. arxiv.org/...2506.09985
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Meta AI (June 11, 2025). "Introducing the V-JEPA 2 world model and new benchmarks for physical reasoning." ai.meta.com/...v-jepa-2-world-model-benchmarks
- ^Meta Newsroom (June 11, 2025). "Our new model helps AI think before it acts." about.fb.com/...odel-helps-ai-think-before-it-acts
- ^1 ^2 ^3 ^4facebookresearch (2025). vjepa2 GitHub repository. github.com/...vjepa2
- Meta AI Research. "Introducing V-JEPA 2." ai.meta.com/...vjepa
- ^1 ^2 ^3LeCun, Y. (2022). *A Path Towards Autonomous Machine Intelligence*. Open Review. openreview.net/forum
- ^Meta AI (June 2023). "I-JEPA: The first AI model based on Yann LeCun's vision for more human-like AI." ai.meta.com/...yann-lecun-ai-model-i-jepa
- ^Meta AI (February 2024). "V-JEPA: The next step toward advanced machine intelligence." ai.meta.com/...t-embedding-predictive-architecture
- ^1 ^2 ^3TechCrunch (June 11, 2025). "Meta's V-JEPA 2 model teaches AI to understand its surroundings." techcrunch.com/...i-to-understand-its-surroundings
- ^1 ^2MarkTechPost (June 12, 2025). "Meta AI Releases V-JEPA 2: Open-Source Self-Supervised World Models for Understanding, Prediction, and Planning." marktechpost.com/...anding-prediction-and-planning
- ^The Robot Report (June 11, 2025). "Meta V-JEPA 2 world model uses raw video to train robots." therobotreport.com/...es-raw-video-to-train-robots
- ^Google (January 29, 2026). "Project Genie: Experimenting with infinite, interactive worlds." blog.google/...project-genie
- ^Mur-Labadia, L. et al. (2026). *V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning*. arXiv:2603.14482. arxiv.org/...2603.14482
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
7 revisions · v7 · 3,258 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent fact-check of the full article against its cited sources (95 claims, 16 sources fetched); 11 material + 16 minor fixed
Cite this page: AI Wiki. "V-JEPA 2." aiwiki.ai, updated 28 Oct 2026, fact-checked 5 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/v_jepa_2

