nanoGPT

Karpathy’s ~300-line GPT trainer — the cleanest way to learn LLM training.

Install Repository Compare

Fit.

Catalog snapshot

Platforms: Linux, Macos, Windows
Operating systems: Linux, macOS, Windows
Ram floor: 16 GB
Gpu: Required · 8 GB VRAM
Docker: No
Setup: Intermediate
Deploy as: Local
Commercial use: Allowed

Education-grade: expect pedagogy, not SOTA checkpoints.

Project.

Fetched Sep 12, 2026¹

License: MIT
Stars: 63k¹
Forks: 10.8k¹
Open issues: 248¹
Language: Python¹
Last commit: Nov 12, 2025¹
Tags: educationgpttraining

About.

nanoGPT is Karpathy’s MIT teaching repo: train a small GPT from scratch in ~300 lines, reproduce the Shakespeare speedrun, or fine-tune GPT-2 class models. The best on-ramp to actually understanding training. Not a product.

GPT training from scratch · char and token models · Shakespeare speedrun · clean Karpathy code

Alternatives.

Nearby.

All Fine-tuning

LllamafileSingle-file LLM runtime from Mozilla. Download, chmod, run.Compare
LLiteLLMProxy that translates many model APIs into one OpenAI-shaped client.Compare
VvLLMHigh-throughput model server for GPU inference.Compare
Lllama.cppC/C++ runtime that made local GGUF models practical.Compare
EexoLink everyday devices into one cluster and run models too big for any single machine.Compare