nanochat Installation

Quick setup for training and inference. nanochat uses uv for fast, reproducible dependency management. Once installed, you can run the speedrun script or any of the individual training and evaluation scripts.

Clone and install

git clone https://github.com/karpathy/nanochat
cd nanochat
uv sync

Requirements

• Python: See .python-version in the repo
• uv: Install uv if needed; see GitHub for setup
• PyTorch: With CUDA for GPU; installs via uv
• Rust: For tokenizer training (optional)

Activate environment

source .venv/bin/activate

On Windows: .venv\Scripts\activate

Verify

Run a quick check to ensure the package loads:

python -c "import nanochat; print('OK')"

Data setup

Before running the full speedrun, you'll need pretraining data. The runs/speedrun.sh script handles data download and preparation. For custom setups, see datasets and the repo's data handling in nanochat/dataset.py.

Next steps

See getting started to run the speedrun and chat UI. Try the hosted demo at nanochat-ai.com.