If you've been using nanoGPT as your go-to minimal GPT implementation for learning and experimentation, you just got a deprecation notice. Andrej Karpathy committed a README update to the nanoGPT repo announcing its deprecation and pointing developers toward a new project: nanochat.

This isn't just a rename. It signals a meaningful shift in how Karpathy thinks about minimal LLM codebases — and what "educational AI code" should look like in 2026.

What was nanoGPT?

Released in late 2022, nanoGPT became the canonical "start here" codebase for anyone who wanted to understand transformer training from scratch. The core train.py fit in ~300 lines. You could train a character-level GPT on Shakespeare in an afternoon on a single GPU.

It filled a gap that no other resource did: it was between "read the Attention Is All You Need paper" and "clone a 100k-line production repo." Thousands of developers used it to build intuition about how pretraining actually works.

What changed — and why nanochat?

The world moved on from pure pretraining. In 2026, the questions developers actually ask are:

  • How does instruction-following get trained into a model?
  • What does RLHF / RLAIF look like in minimal code?
  • How do you build a chat interface that maintains context correctly?

nanoGPT didn't answer any of these — it was designed to train a raw next-token predictor, not a chat assistant. nanochat is built around the chat paradigm from the ground up: system prompts, turn structure, and fine-tuning on conversation data are first-class citizens.

"The README now explicitly tells you: if you want to understand modern LLMs as chat systems, nanoGPT is the wrong starting point."

What this means for developers

If you're learning: Switch your reading to nanochat. The core concepts (attention, positional encoding, layer norm) are the same — but the scaffolding now reflects how models are actually deployed and used today.

If you have existing nanoGPT forks: Your code still works. The repo isn't deleted. But don't expect new issues or PRs to get attention — Karpathy's focus has moved.

If you're building on top of it: nanochat's architecture will be a cleaner base for experimenting with things like DPO, RLHF datasets, and multi-turn context handling — all of which are increasingly relevant if you're building Claude/GPT-based apps and want to understand what's happening under the hood.

The broader pattern

This is part of a consistent pattern in Karpathy's work: he builds the minimal version of whatever the field is currently obsessed with. In 2022 it was "how do you train a GPT." In 2026 it's "how do you build a chat model." Watch nanochat for the same clarity of explanation applied to the alignment-adjacent training loop.

llm.c, his C-based LLM training codebase, is still active and is the place to go if raw performance and low-level understanding is your goal. nanochat sits above it — pedagogically closer to the product layer.

What to do right now

  • Star the nanochat repo and read the README end to end
  • If you have a nanoGPT tutorial or course, update your links
  • Watch Karpathy's X / GitHub for the first nanochat video — it will likely be the "Zero to Hero" entry point for chat model training

The shift from nanoGPT to nanochat is a small commit. But it's Karpathy telling you where the interesting problems are now.