---
title: V-JEPA2-AC-vitg
canonical_url: "https://www.modelscope.cn/models/mlx-community/V-JEPA2-AC-vitg"
md_url: "https://www.modelscope.cn/models/mlx-community/V-JEPA2-AC-vitg.md"
repository: mlx-community/V-JEPA2-AC-vitg
last_updated: 2026-06-06
license: mit
parameters: 1.3B
tensor_type:
  - F16
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 26
stars: 0
tags:
  - mlx
  - video
  - vjepa2
  - world-model
  - action-conditioned
  - robotics
  - apple-silicon
---

# V-JEPA2-AC-vitg

> V-JEPA2-AC-vitg - mlx-community 在 ModelScope 开源的模型。V-JEPA2-AC ViT-g (MLX) — action-conditioned world-model

mlx-community/V-JEPA2-AC-vitg 是 ModelScope 魔搭社区上的 1.3B 参数机器学习模型，采用 mit 许可。

- **Repository**: mlx-community/V-JEPA2-AC-vitg
- **License**: mit
- **Parameters**: 1.3B
- **Tags**: mlx, video, vjepa2, world-model, action-conditioned, robotics, apple-silicon
- **Downloads**: 26
- **Stars**: 0
- **Last updated**: 2026-06-06

Source: https://www.modelscope.cn/models/mlx-community/V-JEPA2-AC-vitg

---

# 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`.