What is V-JEPA 2 and V-JEPA 2.1?
V-JEPA 2 is a family of video models from Meta FAIR, published on 11 June 2025 with weights, code and three new physics benchmarks. It learns by watching: shown a video with parts hidden, it predicts a compressed description of the missing parts rather than the missing pixels. Meta trained it on more than 1 million hours of internet video and 1 million images, with no human labels. The result is an encoder that turns a clip of up to 64 frames into features a small add-on can use to classify actions, anticipate what happens next or answer questions alongside a language model.
That makes it a world model of a different kind from a generative video model. Veo or Genie paint frames you can watch. V-JEPA 2 never draws anything; it predicts in its own internal summary, which is cheaper and ignores detail that does not matter, like the exact shape of leaves. Meta's V-JEPA 2-AC adds a small action-conditioned predictor, trained on under 62 hours of unlabelled robot video from the public DROID dataset, so a robot can score candidate moves against a goal photo and pick the best one.
On 16 March 2026 Meta released V-JEPA 2.1 in the same repository: four new encoders from 80M to 2B parameters at 384-pixel resolution, trained so every patch of the frame carries useful detail, not only the clip as a whole. It is the version to start with for depth, tracking or robot work. The V-JEPA 2 checkpoints (300M to 1B parameters) remain the ones packaged for Hugging Face Transformers. Code is MIT; the Hugging Face cards list MIT or Apache 2.0, and Meta says the models are available for commercial and research use.
Is V-JEPA 2 and V-JEPA 2.1 worth it?
V-JEPA 2 is the strongest openly licensed video encoder for understanding motion, and it plans for a real robot without generating a single frame. Meta's largest V-JEPA 2 model scores 77.3% on Something-Something v2, a test of telling apart actions such as pushing something left or right, against 69.7% for InternVideo2-1B. On Epic-Kitchens-100 action anticipation it reaches 39.7 recall at 5 against 27.6 for the previous best. These figures are Meta's own, measured with a small trained probe on top of the frozen encoder.
The robot results explain why the approach matters. A generative model plans by drawing possible futures, and that is slow: Meta's paper reports NVIDIA's Cosmos taking 4 minutes to choose each action, so one pick-and-place run takes over an hour. V-JEPA 2-AC compares futures in its own compact space and needs 16 seconds per action while sampling ten times as many options. On Franka arms in two labs it never saw in training, it grasped a cup 60% of the time against Cosmos's 0% and Octo's 10%, and completed pick-and-place with a cup 80% of the time. The blog gives 65% to 80% for pick-and-place of new objects with visual sub-goals.
V-JEPA 2.1 raises the ceiling. Its paper reports 40.8 recall at 5 on Epic-Kitchens, 77.7% on Something-Something v2, 7.71 mAP on Ego4D short-term object-interaction anticipation, and a 20-point gain in real-robot grasping over V-JEPA 2-AC. It also produces usable depth maps from a linear probe, 0.307 RMSE on NYUv2, which the first version was not built for. All of this is self-reported, and the 2.1 checkpoints sit on Meta's download server rather than Hugging Face.
The limits are those of a research release. Planning works from a single goal image over roughly 16 seconds of prediction; longer tasks such as pick-and-place need hand-given sub-goals. Meta's paper says results were sensitive to camera position and that the team tried several placements before one worked. On Meta's own physics tests, IntPhys 2, MVPBench and CausalVQA, humans score 85% to 95% while current video models, V-JEPA 2 included, trail well behind, and on IntPhys 2 they sit at or near chance. It reads motion better than the open encoders Meta compares it with; it does not yet understand physics the way a child does.
Where it earns its place
A robotics researcher with a Franka arm and a single RGB camera
Get reach, grasp and pick-and-place working in a new lab without collecting task data
Load V-JEPA 2-AC through PyTorch Hub, give it a goal photo, and let it score candidate moves at each step. Meta reports 60% to 80% success on cup tasks zero-shot; expect to spend time finding a camera position that works, as Meta's own team did.
An ML engineer at a sports-analytics firm with 10,000 labelled clips
Classify fine-grained movements such as dive types or tackles
Freeze a V-JEPA 2 encoder and train only an attentive probe on its features, the method Meta used to reach 90.2% on Diving48. Meta publishes probe checkpoints and configs for SSv2, Diving48 and Epic-Kitchens, so the first baseline is a config change.
A product team at a smart-glasses or home-camera start-up
Anticipate what a person is about to do in egocentric video
Use V-JEPA 2.1, whose paper reports 40.8 recall at 5 on Epic-Kitchens action anticipation, as the encoder, and train a light head on your own footage. The MIT and Apache terms carry none of the use restrictions in Meta's custom DINOv3 and Sapiens2 licences.
Before you commit
- Setup
- pip install the GitHub repo on Python 3.12, or load V-JEPA 2 with Transformers' AutoModel and AutoVideoProcessor; V-JEPA 2.1 and V-JEPA 2-AC load through PyTorch Hub or direct checkpoint downloads. A demo notebook classifies a sample video.
- Integrations
- Hugging Face Transformers (V-JEPA 2 only), PyTorch Hub, timm and einops; probe training runs locally or on SLURM clusters. Meta's paper aligns the encoder with an 8B language model for video question answering.
- Data and privacy
- Self-hosted, so video stays on your hardware. Code is MIT with three files under Apache 2.0; the Hugging Face weights are listed as MIT or Apache 2.0, with no sign-up gate and no use restrictions beyond the licence.
- Cost at scale
- Encoders run from 80M to 2B parameters, so the largest fits on one data-centre GPU; Meta's 16 seconds per planned action is measured on its own hardware. Pre-training from scratch is out of reach for most teams: Meta estimates roughly 60 GPU-years for ViT-g on 64-frame 384-pixel clips.
- Limits
- Clips of up to 64 frames, about 16 seconds of prediction; planning from a single goal image covers short tasks only. Results are self-reported, and on IntPhys 2 current video models including V-JEPA 2 score at or near chance against near-perfect humans.
What’s good
- 77.3% on Something-Something v2 against 69.7% for InternVideo2-1B (Meta)
- Plans robot actions in 16 seconds each against Cosmos's 4 minutes (Meta)
- Permissive licence: MIT code, MIT or Apache 2.0 weights on Hugging Face
- V-JEPA 2.1 adds dense per-patch features for depth and tracking
What’s not
- Outputs features, not answers; every task needs a trained probe or head
- Robot planning needs sub-goals for longer tasks and a carefully placed camera
- V-JEPA 2.1 weights are not on Hugging Face or in Transformers
- Video loader relies on decord, which is unmaintained and does not support macOS
Best for
Robotics and video-ML researchers who need a frozen, permissively licensed video encoder for action recognition, anticipation or planning from goal images.
V-JEPA 2 and V-JEPA 2.1 alternatives
V-JEPA 2 and V-JEPA 2.1: common questions
- Is V-JEPA 2 and V-JEPA 2.1 worth it?
- V-JEPA 2 is the open video encoder to beat for motion and robot planning, but it hands you features and a research codebase, not a finished product. We rate it 4.2 out of 5.
- How much does V-JEPA 2 and V-JEPA 2.1 cost?
- Free. Code under MIT; weights released for commercial and research use, with Hugging Face cards listing MIT or Apache 2.0. No hosted API from Meta.
- What are the downsides of V-JEPA 2 and V-JEPA 2.1?
- Outputs features, not answers; every task needs a trained probe or head; robot planning needs sub-goals for longer tasks and a carefully placed camera; v-JEPA 2.1 weights are not on Hugging Face or in Transformers; video loader relies on decord, which is unmaintained and does not support macOS.
- Who is V-JEPA 2 and V-JEPA 2.1 best for?
- Robotics and video-ML researchers who need a frozen, permissively licensed video encoder for action recognition, anticipation or planning from goal images.
- How much does V-JEPA 2 and V-JEPA 2.1 cost for a team or in production?
- Encoders run from 80M to 2B parameters, so the largest fits on one data-centre GPU; Meta's 16 seconds per planned action is measured on its own hardware. Pre-training from scratch is out of reach for most teams: Meta estimates roughly 60 GPU-years for ViT-g on 64-frame 384-pixel clips.
- Is V-JEPA 2 and V-JEPA 2.1 safe for company data?
- Self-hosted, so video stays on your hardware. Code is MIT with three files under Apache 2.0; the Hugging Face weights are listed as MIT or Apache 2.0, with no sign-up gate and no use restrictions beyond the licence.
- What are the best V-JEPA 2 and V-JEPA 2.1 alternatives?
- The closest alternatives we have reviewed are DINOv3, Genie 4, Molmo2-8B and HY-World 2.0.
Answers checked Nov 9, 2026.
Make V-JEPA 2 and V-JEPA 2.1? Show this rating on your site
