Modelpublic

mlx-community/V-JEPA2-vitl-fpc16-256-ssv2

sourceHugging Facemitupdated 4mo agoView on Hugging Face

1likes11downloads

Model Card

V-JEPA2 ViT-L SSv2 classifier (MLX)

Apple MLX fp16 port of facebook/vjepa2-vitl-fpc16-256-ssv2 — V-JEPA2 ViT-L encoder + attentive-pooler + linear head for Something-Something-v2 action recognition (174 classes). MIT.

bash

pip install vjepa2-mlx   # https://github.com/xocialize/vjepa2-mlx
vjepa2-mlx -i clip.mp4 --task classify

—Parity vs PyTorch (cpu fp32): logits rel 2.43e-6, argmax matches (fp16 too).
—Precision: fp16 (~707 MB).