nanochat is Andrej Karpathy’s open-source project — nearly 59,000 stars on GitHub — that builds a complete ChatGPT-style system on a single GPU node for roughly $100 of compute, billed as “the best ChatGPT that $100 can buy.”
Where nanoGPT taught the model, nanochat teaches the whole product: a single, hackable codebase that covers tokenization, pretraining, mid-training, supervised fine-tuning, reinforcement learning, evaluation and finally inference behind a chat web UI. Everything runs on one 8-GPU node, and the default speedrun recipe targets about $100 of cloud compute (roughly a few hours), scaling to stronger models if you spend more. The code is deliberately dependency-light and readable end to end, so a student can trace exactly how a raw text corpus becomes a conversational assistant. Released under the MIT license, the project has trended 13 times in the archive and was actively developed through September 2026.
nanochat is intentionally the opposite of a framework: there is almost nothing between you and the machinery, so every design decision — from the Rust-based tokenizer to the RL reward — is inspectable. Karpathy maintains it as a living reference that improves as better recipes are discovered, and its leaderboard-style baselines give learners objective targets at each budget level.
nanochat is free and open source under the MIT license. The cost is your compute — the flagship recipe is designed to cost about $100 on rented GPUs.
Advanced students, ML engineers and the curious who want to build a working ChatGPT clone end to end and understand every stage, not just fine-tune a checkpoint.