Release V-JEPA 2.1 on Hugging Face

[NielsRogge]

Hi @Adrien987k 🤗

I'm Niels and work as part of the open-source team at Hugging Face. I discovered your work through Hugging Face's daily papers as yours got featured: https://huggingface.co/papers/2603.14482.
The paper page lets people discuss about your paper and lets them find artifacts about it (your models for instance), you can also claim the paper as yours which will show up on your public profile at HF, add Github and project page URLs.

I saw that you've already hosted the previous V-JEPA 2 models on the hub (under the facebook organization), which is great! Would you like to also host the new V-JEPA 2.1 pre-trained checkpoints (the ViT-B, L, g, and G variants) as well as the V-JEPA 2-AC model on https://huggingface.co/models?

Hosting the new weights on Hugging Face will give them more visibility and enable better discoverability through metadata tags. It also allows the community to load them easily via hf_hub_download or potentially directly through integration with the transformers library, which you've already implemented for the V2 models.

If you're down, I can help with the process. You can use the PyTorchModelHubMixin class for custom models, or simply upload the .pt files to a new repository.

After uploaded, we can also link the models to the paper page (read here) so people can discover your latest work.

Let me know if you're interested or need any guidance!

Kind regards,

Niels

[coder0143]

Hello @NielsRogge , thanks for your amazing work on hf open source models,
I'm currently a core contributor to google's jax-ml/bonsai , I have previously integrated the vjepa2 model with jit support here: https://github.com/jax-ml/bonsai/tree/main/bonsai/models/vjepa2 , built based on the hf implementation of the model following Aritra Roy Gosthipaty's work.
Thanks for opening this issue, once the models are added to hf, I would like to extend support to jax-ml as well.
Thankyou, I have also sent a connection request as well on linkedin! My profile: Pratham Shah

[davevanveen]

Thanks for kicking this off @NielsRogge, and great to see the community interest @coder0143!

I've opened a PR on huggingface/transformers adding V-JEPA 2.1 inference support to the existing vjepa2 model family: huggingface/transformers#45497

It extends VJEPA2Config and VJEPA2Model with backward-compatible support for all four 2.1 checkpoints (ViT-B/L/g/G at 384), including the conversion script. Verified against Meta's reference implementation on both ViT-B and ViT-g.

Once the 2.1 weights are uploaded to the Hub, users will be able to load them directly via from_pretrained.

[coder0143]

@davevanveen Thanks for the pr to hf!

[apiantonio]

A data point in support of this request, plus an interim resource for anyone landing here.

I converted all four V-JEPA 2.1 encoders (ViT-B/L/g/G @ 384px) to HF format and validated
them against the official torch.hub implementation:

variant	weights vs checkpoint	forward parity, T=16 / 32 / 64
ViT-B	bit-exact, 0 orphaned params	0 / 0 / 0
ViT-L	bit-exact, 0 orphaned params	5.85e-04 / 1.00e-03 / 0
ViT-g	bit-exact, 0 orphaned params	0 / 0 / 0
ViT-G	bit-exact, 0 orphaned params	0 / 0 / 0

fp32, same input tensor through both implementations. The ViT-L residual is GEMM
factorization (fused vs split QKV changes accumulation order) reproduced by a control
experiment to four significant figures, not a conversion error.

Models: https://huggingface.co/collections/apiantonio/v-jepa-21-huggingface-ports
Conversion tooling and modeling code: https://github.com/Dev-Jahn/vjepa2-hf

To be clear, this doesn't substitute for what's being asked here. These are community
trust_remote_code ports; official checkpoints under the facebook org plus native
transformers support would be strictly better.

[davevanveen]

Thank you for the models and detailed analysis, @apiantonio! Really helpful to bridge the gap.

Working with @molbap to close out the official PR #45497 shortly, at which point native Transformers support and checkpoints under the Facebook org will be available.