nanochat

nanochat is a 561M parameter transformer language model trained for conversational AI tasks. This model demonstrates that capable chat models can be trained efficiently on modest hardware budgets (~$100 on 8x H100 GPUs).

Read about the process at https://samdobson.uk/posts/training-a-chatgpt-clone-for-cheap/

Chat with the model at https://huggingface.co/spaces/sdobson/nanochat

Model Description

• Developed by: Andrej Karpathy
• Trained by: Sam Dobson
• Model type: Transformer-based causal language model
• Language(s): English
• License: MIT
• Parameters: 560,988,160 (~561M)

Architecture

• Layers: 20
• Hidden size: 1280 channels
• Attention heads: 10
• Head dimension: 128
• Vocabulary size: 65,536 tokens

Training Details

Training Data

nanochat was trained in multiple stages:

1. Pretraining: 100B token subset of FineWeb-EDU (11.2B tokens processed)
2. Midtraining: SmolTalk conversations, MMLU multiple choice questions, GSM8K math problems
3. Supervised Fine-tuning (SFT): Conversational adaptation data

Training Procedure

Tokenization

• Custom Rust-based tokenizer
• Vocabulary: 65,536 tokens
• Compression ratio: 4.8 characters per token

Training Infrastructure

• Hardware: 8x H100 GPUs (Lambda GPU Cloud)
• Training time: ~3 hours for pretraining stage
• Estimated compute: ~4e19 FLOPs
• Total cost: ~$100

Training Stages

The model was trained in three stages:

1. Pretraining on web text (FineWeb-EDU)
2. Midtraining on domain-specific datasets (reasoning, conversation, maths)
3. Supervised fine-tuning for chat optimisation

Performance

Benchmark Results

Benchmark	Score	Description
MMLU	23.99%	Multitask language understanding
GSM8K	4.47%	Grade school math problems
HumanEval	6.71%	Python code generation
ARC-Easy	24.79%	Science questions (easy)
ARC-Challenge	24.32%	Science questions (hard)
ChatCORE	1.73%	Conversational reasoning

Intended Use

Direct Use

nanochat is designed for:

• Conversational AI applications
• Research on efficient language model training
• Educational purposes for understanding LLM training pipelines
• Low-resource deployment scenarios

Downstream Use

The model can be fine-tuned for specific conversational tasks or used as a base model for further domain adaptation.

Out-of-Scope Use

• Production-grade conversational AI (the model is relatively small and has limited capabilities)
• Tasks requiring specialised knowledge or high accuracy
• Critical applications where errors could cause harm

Limitations and Bias

• Small scale: At 561M parameters, this model has significantly fewer capabilities than larger models (1B+ parameters)
• Limited training: Trained on only 11.2B tokens, which is modest by modern standards
• Performance: Benchmark scores indicate limited reasoning and mathematical capabilities
• Bias: Inherits biases from training data (FineWeb-EDU, SmolTalk, etc.)
• Language: English-only

Inference guide

Simon Willison created a script to allow this to run on CPU on MacOS:

  cd /tmp
  git clone https://huggingface.co/sdobson/nanochat
  uv run https://gist.githubusercontent.com/simonw/912623bf00d6c13cc0211508969a100a/raw/80f79c6a6f1e1b5d4485368ef3ddafa5ce853131/generate_cpu.py \
    --model-dir /tmp/nanochat \
    --prompt "Tell me about dogs."

Otherwise you can:

1. Download all files
2. Put tokenizer.pkl and token_bytes.pt in ~/.cache/nanochat/tokenizer
3. Put model_000650.pt and meta_000650.json in ~/.cache/nanochat/chatsft_checkpoints/d20
4. Clone https://github.com/karpathy/nanochat
5. Run uv sync followed by uv run python -m scripts.chat_web

Citation

Repository: github.com/karpathy/nanochat

@software{nanochat2025,
  author = {Karpathy, Andrej},
  title = {nanochat: A 561M parameter conversational language model},
  year = {2025},
  url = {https://github.com/karpathy/nanochat}
}

Model Card Author

Sam Dobson