TopGit
GitHub Repo Review

nanoGPT review: a simple GPT trainer, now deprecated

karpathy/nanoGPT
NTopGit review image for karpathy/nanoGPT
Review by Topgit.dev for karpathy/nanoGPT, with GitHub repository stats and README context.
Quick verdict

nanoGPT is Andrej Karpathy's minimal, from-scratch PyTorch codebase for training and finetuning GPT-2-scale language models, and its own README now labels it old and deprecated, pointing to a newer project called nanochat instead. Reach for nanoGPT if you want to read and modify a complete GPT training loop in a few hundred lines; skip it if you need an actively maintained project, since nanochat is what the author now recommends.

Stars
★ 64.7k
Forks
⑂ 11.1k
Language
Python
License
MIT
Topic
—
Updated
Nov 2025
Homepage
—
GitHub

The nanoGPT README now opens with a notice that the repository is old and deprecated, kept online for posterity, and that most visitors probably wanted its successor, nanochat. Below that notice the pitch is unchanged: the simplest, fastest code for training and finetuning medium-sized GPTs, plus a Shakespeare quick start that runs on a laptop. Here is how each promise holds up against what people wrote on Hacker News and in GitHub issues.

Simplicity matches best. On Hacker News, @beardyw, 2025-10-10 described it as "reasonably accessible and easy to run". @efm, 2025-01-30 framed the purpose the same way: "The use case for this is learning in simple example." That is a learning use, and says nothing about speed.

The laptop promise is also supported, with a caveat about quality. @fzzzy, 2025-01-30 argued that "Learning how to train a 124M is just as useful as a 700B, and is possible on a laptop." @faxmeyourcode, 2025-10-13 started on it over a weekend, called it "an older but fabulous learning exercise", and admitted the tiny Shakespeare model is poor: "Results are about what you'd expect from that, they suck".

Speed is where the README and the evidence diverge. The headline training time assumes the 8XA100 40GB node, and @lostmsu, 2026-09-30 noted that the code is optimized but tuned for that setup, so scaling experiments at home means going faster. In an issue, @JustinKunzi, 2023-06-17 wrote that a 4x NVIDIA A4500 machine "had no problem running the Shakespeare training set and the outcome was as intended." The collected text breaks off before saying how their GPT 124M run on the OWT dataset ended, so the full-size claim has no confirmation here. In another issue, @holyseven, 2023-02-20 reported NaN loss after about 6000 iterations, even after lowering the learning rate to 1e-5, and wondered whether the float16 dtype in their config was to blame.

The numbers set the scale. The 2022-12-28 creation date sits against 64,671 stars and 11,130 forks, yet the evidence for this review is thin: a handful of Hacker News comments, most dated 2025 or earlier, and four GitHub issues from 2023. Nobody in them describes a recent run.

What the README leaves out is what the deprecation means in practice. Its last push was on 2025-11-12, and 358 issues are still open. The README does not say whether questions like the NaN report above will get answers, or which of the old scripts nanochat replaces.

View on GitHub ↗
Source & attribution

Facts and quotes sourced from the karpathy/nanoGPT GitHub repository and its README.

GitHub data · last synced Nov 12, 2026Reviewed by Henry
← Back to TopGit

Is nanoGPT worth your time?

ChatGPT, Claude and Perplexity can all read this page. Ask one of them what it makes of nanoGPT.

GitHub