nanochat

For roughly the price of a dinner out, nanochat trains a working ChatGPT-style assistant from scratch: Andrej Karpathy's follow-up to nanoGPT covers the entire stack in one clean codebase, from tokenizer training through pretraining, supervised fine-tuning, optional reinforcement learning, evaluation, and an inference server with a web chat UI. The reference speedrun script targets an 8xH100 node and reaches GPT-2 grade capability in roughly 1.5 hours for about 48 dollars, tracked on a time-to-GPT-2 leaderboard in the repository, while the stated goal is the best models reachable end to end on budgets under 1,000 dollars. Setup uses the uv package manager with CUDA as the primary path plus CPU and Apple Silicon fallbacks for tinkering, though real training assumes a multi-GPU node rented by the hour. The code is MIT licensed and deliberately minimal, hackable, and dependency-light. Released in October 2025, it collected more than 56,000 GitHub stars within months and is planned as the capstone project of Karpathy's LLM101n course, making it the standard answer to how to train your own chatbot end to end.