---
license: mit
library_name: mlx
pipeline_tag: video-classification
tags: [mlx, video, vjepa2, world-model, action-conditioned, robotics, apple-silicon]
---

# V-JEPA2-AC ViT-g (MLX) — action-conditioned world-model

Apple **MLX** fp16 port of Meta's **V-JEPA2-AC** (`vjepa2-ac-vitg`): a **ViT-g**
video encoder + an **action-conditioned predictor** that, given encoder context
tokens + per-frame 7-DoF robot poses (action/state), predicts future latent
states — the world-model used for robot planning. MIT.

```python
from vjepa2_mlx.utils.weights import build_ac_encoder, build_ac_predictor
enc = build_ac_encoder()          # ViT-g encoder (hidden 1408, 40 layers)
pred = build_ac_predictor()       # AC predictor (frame-causal, 3D-RoPE)
# tokens = enc(video); future = pred(tokens, actions, states)
```

- **Arch**: ViT-g encoder (1408 / 40 / 22 heads, GELU, 3D-RoPE) + AC predictor
  (1024 / 24 / 16, fused-qkv ACRoPEAttention, frame-causal mask, 7-DoF action+state encoders).
- **Parity vs upstream torch (cpu fp32)**: encoder rel 3.5e-5 · AC predictor rel 8.8e-6
  (structural 8.9e-8). fp16: encoder rel 9.1e-3 · predictor 4.8e-4.
- **Precision**: fp16 (~2.6 GB).

MIT (© Meta Platforms). Converted from `vjepa2-ac-vitg.pt`.