dgrauet/vjepa-2.0-vitl-mlx

MLX format conversion of facebook/vjepa2-vitl-fpc64-256.

Converted with mlx-forge.

Usage

These weights can be used with vjepa2-mlx:

git clone https://github.com/dgrauet/vjepa2-mlx.git && cd vjepa2-mlx
uv pip install -e packages/vjepa2-core-mlx

python - <<'PY'
from vjepa2_core_mlx.utils.weights import from_pretrained

encoder = from_pretrained("dgrauet/vjepa-2.0-vitl-mlx")   # or a local converted directory
feats = encoder(clip)                    # clip: B, C, T, H, W (channels-first)
PY

Related Projects

• Upstream (Meta): https://github.com/facebookresearch/vjepa2
• MLX inference port: https://github.com/dgrauet/vjepa2-mlx

Files

• config.json (833.00 B)
• diving48_probe.safetensors (188.40 MB)
• ek100_probe.safetensors (203.68 MB)
• encoder.safetensors (1.13 GB)
• predictor.safetensors (84.27 MB)
• split_model.json (1008.00 B)
• ssv2_probe.safetensors (188.89 MB)