Modelpublic

mlx-community/V-JEPA2-vitl-fpc64-256

sourceHugging Facemitupdated 4mo agoView on Hugging Face

1likes17downloads

Model Card

V-JEPA2 ViT-L (MLX) — encoder + JEPA predictor

Apple MLX fp16 port of Meta's V-JEPA2 ViT-L/16 (facebook/vjepa2-vitl-fpc64-256): video/image embedding extraction plus the JEPA latent-space predictor (masked world-model). MIT.

bash

pip install vjepa2-mlx   # https://github.com/xocialize/vjepa2-mlx
vjepa2-mlx -i clip.mp4 --task embed -o emb.npy

python

from vjepa2_mlx.pipeline_mlx import embed_video
emb = embed_video("clip.mp4", num_frames=16)   # (1024,)

—Arch: ViT-L/16, 24 layers, hidden 1024, 16 heads, 3D-tubelet Conv3d patch embed, 3D-RoPE; predictor 384/12/12.
—Parity vs PyTorch (cpu fp32): encoder rel 2.66e-5 · predictor rel 1.67e-6.
—Precision: fp16 (~650 MB; encoder fp16 rel 4.7e-3).