V-JEPA 2.1 ViT-G/16 (384 px)
Release in the V-JEPA 2 family · version 2.1 ViT-G/16, 384 resolution (vjepa2_1_vit_gigantic_384)
Maintained by Meta (FAIR)16
The largest V-JEPA 2.1 model, a 2-billion-parameter ViT-G/16 video and image encoder released by Meta in March 2026. V-JEPA 2.1 changes the V-JEPA 2 recipe to learn dense, temporally consistent features, using a dense predictive loss over all tokens, self-supervision at several intermediate layers, and separate tokenizers for images and videos. Meta distilled this model into smaller ViT-B and ViT-L variants.16
- Repository: vjepa2 repository (V-JEPA 2.1 checkpoints and code) (external site: github.com)
- Paper: V-JEPA 2.1 paper (arXiv 2603.14482) (external site: arxiv.org)
- License: LICENSE (MIT, repository) (external site: github.com)
Availability and license
Overall availability
Direct checkpoint download linked from the repository README, or loaded through PyTorch Hub as vjepa2_1_vit_gigantic_384. This review found no Hugging Face repository for V-JEPA 2.1.1
Availability is separate from permission: read the license before using or redistributing.
The repository README says most of the project is licensed under MIT, with three data-augmentation and worker-initialization files under Apache 2.0. It does not name a separate license for the V-JEPA 2.1 checkpoints, so no weights license is recorded here.1
Model-disclosure tier
The model parameters for this release can be downloaded by the public. License terms may still restrict use, redistribution, or commercial use.
No license for the weights is recorded in this catalog.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Public | Direct download link in the README; no gating.1 |
| Inference codeIs code for running the model published? | Public | The repository provides PyTorch Hub loaders for the V-JEPA 2.1 models.1 |
| Training codeIs the code used to train the model published? | Public | The repository includes a V-JEPA 2.1 pretraining loop (app/vjepa_2_1) and pretraining and cooldown configurations for ViT-G/16.13 |
| Training-data informationPublic = the training data itself can be obtained. Partial = composition or sources are documented without full access. | Partial | The paper describes VisionMix-163M, which replaces V-JEPA 2's ImageNet images with LVD-142M and rebalances the video sources (SSv2, Kinetics, HowTo100M, YT-Temporal-1B). LVD-142M is a curated image set from Meta's earlier work that this review did not find released.6 |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Partial | The paper describes the training phases (135,000 pretraining iterations, then a 12,000- iteration cooldown at higher resolution and more frames), and configs are published. The published ViT-G/16 cooldown config is named for 256-pixel input, while the checkpoint is a 384-pixel model; this review did not confirm that the configs reproduce the released checkpoint.631 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Public | The repository publishes frozen-evaluation code and a V-JEPA 2.1 evaluation config set for this model (configs/eval_2_1/vitG-384), covering SSv2, Diving48, EPIC-KITCHENS-100, Kinetics-400, ImageNet-1k, COIN, and Jester. This review did not confirm configs for every task reported in the paper, such as depth estimation or robot grasping.14 |
What it is useful for
Organization context
Other releases in the V-JEPA 2 family
- V-JEPA 2 ViT-g/16 (384 px)Model-disclosure tier (USASI rubric v0.1): Open-stack
U.S. eligibility
Eligible · basis: U.S. headquarters
The paper lists all authors as affiliated with FAIR at Meta (the first author also lists Universidad de Zaragoza, with the work done at Meta), and the checkpoint and code are published by Meta in the facebookresearch vjepa2 repository under a Meta Platforms copyright. Meta Platforms, Inc. has its principal executive offices in Menlo Park, California, per its Form 10-K.6127
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.