Most people use AI models. Far fewer ever build one from scratch. nanochat, by Andrej Karpathy, exists to close that gap: a minimal, hackable harness that takes you through the whole journey of training a ChatGPT-style model end to end — on a single GPU node.
nanochat (karpathy/nanochat) was first published on 13 October 2025. It is written in Python with PyTorch and released under the MIT licence. Its tagline is "The best ChatGPT that $100 can buy." — but note that "$100" is budget framing: the README's worked figure for a GPT-2-grade model is about $48, roughly two hours on 8×H100 GPUs at about $3 per GPU-hour, or about $15 on spot pricing.
Key facts
| Item | Details |
|---|---|
| Repository | karpathy/nanochat |
| Author | Andrej Karpathy |
| What it is | Minimal, hackable harness for training a ChatGPT-style model from scratch |
| Pipeline | Tokenizer → pretraining → SFT → evaluation → inference |
| Built with | Python / PyTorch |
| Licence | MIT |
| First published | 13 October 2025 |
| Tagline | "The best ChatGPT that $100 can buy." |
| GPT-2-grade model cost | ~$48 (~2 hours on 8×H100 at ~$3/GPU-hour); ~$15 on spot |
| GPT-2 speedrun record | 1.65 hours (14 March 2026) |
| GitHub stars | 60,705 / 8,524 forks as of 19 November 2026 (RepoRadar) |
The star count stood at 59,646 with 8,388 forks on 2 November 2026, and reached 60,705 stars with 8,524 forks as of 19 November 2026. Star counts measure interest, not production use — they show how many developers are watching a project, not how many are running it in production.
The repository has been quiet lately: its last push was on 7 September 2026, and the latest leaderboard record dates to 14 March 2026.
What this means
nanochat matters most as a learning tool. Training a language model is usually presented as something only large labs can do; nanochat shows the full pipeline — tokenizer, pretraining, supervised fine-tuning, evaluation and inference — in a codebase small enough for one person to read and change. The GPT-2 speedrun record of 1.65 hours, set on 14 March 2026, shows how far that compact approach can be pushed.
For students and engineers, the honest takeaway is about understanding, not shortcuts: nanochat teaches how a ChatGPT-style model is actually built, at a budget a determined learner can afford. It does not turn a small model into a frontier one — and the quiet repository since September suggests the project has settled into a stable, finished shape rather than active expansion.
A note on verification
The facts in this post come from the verified research record for this item, including the repository details and the date-stamped star counts listed above. Star counts change daily, so the figure quoted here is a snapshot as of 19 November 2026, not a live count. Cost figures are the project's own worked estimates.
Source: karpathy/nanochat repository record; star counts via RepoRadar. Last checked: 21 November 2026.
FAQs
What is nanochat?
nanochat is Andrej Karpathy's minimal, hackable harness for training a ChatGPT-style model from scratch end-to-end on a single GPU node — covering the tokenizer, pretraining, supervised fine-tuning, evaluation and inference.
How much does it cost to train a model with nanochat?
The project's tagline says "The best ChatGPT that $100 can buy", but the README's worked figure for a GPT-2-grade model is about $48 — roughly two hours on 8×H100 GPUs at about $3 per GPU-hour, or about $15 on spot pricing.
Is nanochat still actively developed?
The repository has been quiet: its last push was on 7 September 2026 and the latest leaderboard record dates to 14 March 2026, when the GPT-2 speedrun record of 1.65 hours was set.
How popular is nanochat on GitHub?
As of 19 November 2026, the karpathy/nanochat repository had 60,705 GitHub stars and 8,524 forks. Star counts measure interest, not production use.
Comments (0)
Share your question or experience about this syllabus. Your comment will appear after review.
No comments yet — be the first to share your thoughts!