---
license: mit
library_name: pytorch
tags:
  - rubiks-cube
  - reinforcement-learning
  - deepcubea
  - pytorch
pipeline_tag: other
---

# DeepCube — Cube3 (3×3×3) cost-to-go network

PyTorch weights for a DeepCubeA-style cost-to-go network that solves the 3×3×3 Rubik's cube via weighted A\*.

- **Input**: one-hot encoded cube state (324 dims = 54 stickers × 6 colors)
- **Output**: scalar cost-to-go estimate (predicted moves to solved state)
- **Architecture**: MLP — see `deepcube/model.py` in the source repo
- **Training**: Approximate Value Iteration on random scrambles (see `train.ipynb`)
- **Source code**: https://github.com/ac1982/deepcube

## Files

- `deepcube_cube3.pt` — final weights (keys: `net`, `cfg`, `loss_hist`, `elapsed`)

## Usage

```python
from huggingface_hub import hf_hub_download
from deepcube.model import load_checkpoint

ckpt_path = hf_hub_download("alexever/deepcube-cube3", "deepcube_cube3.pt")
model, cfg = load_checkpoint(ckpt_path)
```

Or via the bundled server, which auto-loads from `checkpoints/deepcube_cube3.pt`:

```bash
huggingface-cli download alexever/deepcube-cube3 deepcube_cube3.pt \
    --local-dir checkpoints
python -m deepcube.server
```

## License

MIT — see source repository.