Skip to content

nanochat

Train a ChatGPT-like model end to end on one node

Visit official site

What nanochat is

The project is described as a full-stack implementation of a ChatGPT-like large language model that runs on a single 8XH100 node from training through inference. The documented total cost is about 100 USD, with a simple code structure meant to be easy to understand and modify.

What you can do with it

  • Run the full pipeline from data preparation to inference
  • Talk to your own model through a simple ChatGPT-like interface
  • Train in about 4 hours with the speedrun.sh script
  • Modify the code or scale up to larger models

Application scenarios

  • Educational institutions build learning support tools
  • Small startups create customer service chatbots
  • Developers run research experiments on language models

Getting started

  • Prepare an 8XH100 GPU node with the required dependencies
  • Download and clone the project to a local machine
  • Activate the Python virtual environment
  • Run the speedrun.sh script to start training

Pros & cons

✓ What we like

  • The whole pipeline runs on a single 8XH100 node
  • The code is described as easy to modify and customise

! What to watch out for

  • The described setup centres on a single 8XH100 GPU node
  • The about-4-hour figure is tied to the speedrun.sh script

FAQ

What hardware is described?

The material names a single 8XH100 GPU node for training and inference.

How long does training take?

The speedrun.sh script is stated to finish training in about 4 hours.

Last reviewed: 2026-11-15

More AI chatbot tools

View all →

How we review