V-JEPA 2

Meta, United States 路 June 2025

Meta FAIR

World model 路 Open weights, MIT 路 Also written V-JEPA 2 ViT-g, vjepa2-vitg-fpc64-256

Parameters
1 B
1,034,555,264 parameters for the ViT-g checkpoint as counted on Hugging Face.
GPU memory
1.9 GB
weights at bf16, computed. Fits an 8 GB card
Inference latency
not published
per action step
Weights
MIT
MIT.

What it is

V-JEPA 2 is Meta's self-supervised video model and the base of its world model work. It learns by predicting masked parts of a video in a learned feature space instead of in pixels, trained on more than 1 million hours of internet video, and is released as encoders from ViT-L (300M) to ViT-g (1B). On its own it is a perception model: 77.3 percent top-1 on Something-Something v2 and 39.7 recall at 5 on Epic-Kitchens-100 action anticipation. For robots it matters as the frozen encoder under V-JEPA 2-AC, the action-conditioned version that plans pick-and-place on Franka arms. V-JEPA 2.1, released in March 2026, adds checkpoints from 80M to 2B with better dense features.

Architecture

Backbone
Vision transformer video encoder (ViT-L/16, ViT-H/16 or ViT-g/16) with a predictor, pretrained by masked prediction of latent video features. Not a vision language model.
Action head
None. The pretrained model outputs video representations; actions only enter in the post-trained V-JEPA 2-AC.
Parameters
1,034,555,264 parameters for the ViT-g checkpoint as counted on Hugging Face. The README lists ViT-L/16 at 300M, ViT-H/16 at 600M and ViT-g/16 at 1B, and for V-JEPA 2.1 ViT-B at 80M up to ViT-G at 2B.
Pretraining data
More than 1 million hours of internet video, according to the paper's abstract.
Embodiments
None directly; robot use goes through V-JEPA 2-AC on Franka arms

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 1 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp323.9 GBTraining master weights, rarely used for inference
bf16 or fp161.9 GBThe usual way these checkpoints are served
int81.0 GBQuantised, expect some loss of precision on fine motions
int40.5 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

No memory or latency figure is published.

The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

No results in this category are published for this model.

Real world

No results in this category are published for this model.

Fine tuned tasks

No results in this category are published for this model.

On real hardware

None for the encoder itself. Robot results belong to V-JEPA 2-AC, which keeps this encoder frozen.

Fine tuning it yourself

Weights load through Hugging Face Transformers (AutoModel) or PyTorch Hub, and attentive probes for Something-Something v2, Diving48 and Epic-Kitchens-100 are published. For robot planning, the action-conditioned predictor is post-trained on top of the frozen ViT-g encoder with the DROID config in the repository.

Where it helps, where it does not

Strengths

  • Open checkpoints at several sizes under permissive licences.
  • Strong motion understanding: 77.3 percent top-1 on Something-Something v2 and 39.7 recall at 5 on Epic-Kitchens-100.
  • Predicts in representation space rather than in pixels, the design that V-JEPA 2-AC builds on for planning.

Limits

  • Not a policy and not a video generator: on its own it neither outputs actions nor renders frames.
  • Its headline numbers are video understanding benchmarks, not robot results.
  • Licence tags differ between the repository (MIT) and some Hugging Face checkpoints (Apache-2.0); check the one you download.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-11-12.

Try a policy on a real arm

A real SO-100 runs on /live and can be driven from the browser when it is online. Five of the models in this arena can be fine tuned on a dataset you record with your own arm.