申請取用 · Request access

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

逐個人工審批，請說明單位與用途。本倉庫是 Meta V-JEPA 2.1 官方權重的格式轉換產物
（MIT，保留 Meta Platforms 版權聲明），不是新模型，亦不含任何視頻或幀。
加載方法、數值一致性核對與部署注意事項隨倉庫文件提供，審核通過後可見。

Manual review — please state your affiliation and intended use. This repository is a
format conversion of Meta's official V-JEPA 2.1 weights (MIT, Meta Platforms
copyright retained); it is not a new model and contains no video or frames.
Loading instructions, numerical checks and deployment notes ship with the repository
files and are visible once approved.

V-JEPA 2.1 ViT-B (384) — encoder weights for Jetson and Ascend 910C

English · 简体中文 · 繁體中文

Encoder-only FP16 weights of Meta's V-JEPA 2.1 ViT-B/16 (384), packaged so the same checkpoint runs on an NVIDIA Jetson Orin NX 8GB edge GPU and a Huawei Ascend 910C NPU — the same weights, the same input, the same script on both, with outputs verified against a CPU FP32 reference.

166 MB instead of the original 1.66 GB. This is a deployment artifact, not a new model: the weights are a format conversion of the official release.

Measured performance

FP16, batch 1, 16 frames × 384 × 384:

Device	Latency P50	Throughput	Peak device memory
Jetson Orin NX 8GB	308 ms	3.2 clips/s	341 MB
Ascend 910C (one die)	65.6 ms	15.2 clips/s	396 MB

Running cleanly on Ascend matters: most published vision checkpoints have never been validated on that stack at all.

What you get

The FP16 encoder checkpoint, a fixed-input specification, and a CPU FP32 reference output so you can verify your own deployment. Loading instructions and deployment notes are included for approved users.

Intended use

Video feature extraction on edge GPUs and Ascend NPUs; a portability and benchmarking reference. Not validated for any safety-critical or surveillance decision-making.

Licence and credit

Licence: MIT — inherited from upstream; the Meta Platforms copyright notice is retained in LICENSE.
Upstream: Meta's V-JEPA 2.1, source checkpoint vjepa2_1_vitb_dist_vitG_384.pt (key ema_encoder) from facebookresearch/vjepa2. The weights are not retrained here.

@misc{vjepa21,
  title         = {V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning},
  author        = {{Meta FAIR}},
  year          = {2026},
  eprint        = {2603.14482},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2603.14482}
}

Citation

If you use this repository, please cite it. The upstream works credited above should be cited as well.

@misc{lin2026vjepaascend,
  author       = {Zeteng Lin and Jing Tang},
  title        = {V-JEPA 2.1 ViT-B (384) encoder weights for Jetson and Ascend 910C},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Zeteng/V-JEPA2.1-Ascend}}
}

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

Video Classification

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Paper for Zeteng/V-JEPA2.1-Ascend

V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning

Paper • 2603.14482 • Published Mar 15 • 39