⬢github Python · 60.3K ★ +671 since we first saw it · pushed 58 d ago · MIT
karpathy/nanochat
The best ChatGPT that $100 can buy.
nanochat is a minimal, hackable Python harness for training LLMs end-to-end on a single GPU node, covering tokenization, pretraining, finetuning, evaluation, and inference. It trains GPT-2-capability models for tens of dollars with a single `--depth` knob and includes a 'time-to-GPT-2' speedrun leaderboard.
Why now: The repo maintains an actively-updated speedrun leaderboard pushing 'time to GPT-2' from 168 hours (2019) down to under 2 hours, plus ongoing autoresearch experiments, driving community interest and collaboration.
Who it is for: ML researchers, students, and developers who want to learn or experiment with the full LLM training pipeline on modest hardware.
Stars over our 243 snapshots: 59.6K to 60.3K, since 2 d ago.
Where people talked about it
- ⬢github karpathy/nanochat 1 d ago
API: https://socialmediatrends-api.osmike.com/v1/repos/karpathy/nanochat