README.md

---
license: other
tags:
  - mlx
  - robotics
  - pi0.5
  - quantized
  - apple-silicon
---

# pi0.5 — 4-bit Quantized MLX Weights

4-bit quantized weights for [lerobot/pi05_base](https://huggingface.co/lerobot/pi05_base) converted to Apple MLX format.

Runs on **Apple Silicon (M1/M2/M3)** with ~2.4 GB RAM. Loads in ~6s, inference in ~2s per action chunk.

## Architecture
- PaliGemma 2B VLM (SigLIP + Gemma 2B) + Gemma 300M action expert
- Flow-matching policy: 10-step Forward Euler denoising
- Output: action chunk [B, 50, 32]

## Usage

```python
from huggingface_hub import hf_hub_download
import mlx.core as mx
import mlx.nn as nn

# Download quantized weights (~2.6 GB, one-time)
npz_path = hf_hub_download("mohan007/pi05-mlx-4bit", "pi05_mlx_4bit.npz")

# Load with mlx_pi05
from mlx_pi05.load import load_model
model = load_model(quantized_path=npz_path, quantize=True)
model.eval()

# Run inference
import numpy as np
image_mlx = mx.array(np.zeros((1, 3, 224, 224), dtype=np.float32))
lang_mlx  = mx.array(np.array([[1, 2, 3, 4, 5]], dtype=np.int32))
actions   = model.sample_actions(image_mlx, lang_mlx)  # [1, 50, 32]
```

## Quantization
- Gemma 2B + expert layers: 4-bit (group_size=64)
- SigLIP kept in float16 (fc2 input dim 4304 not divisible by 64)
- Total: ~2.4 GB vs ~7.2 GB float16

## Source
Converted from [lerobot/pi05_base](https://huggingface.co/lerobot/pi05_base) original float32 safetensors.