Redazione RHC : 14 October 2025 09:38
Developer Andrej Karpathy has unveiled nanochat , a minimalist, fully open-source version of ChatGPT that can be trained and run on a single computer. Designed as a learning platform for Eureka Labs’ LLM101n course , the project allows users to build their own language model “from scratch to the web interface” without cumbersome dependencies or complex infrastructure.
The goal of nanochat is to demonstrate that a basic analog of ChatGPT can be built in a few hours and for about $100. The speedrun.sh script automatically performs all the steps, from tokenization and training to inference and launching a web interface that can be used for communication, just like ChatGPT.
On a node with eight NVIDIA H100 GPUs , the entire process takes about four hours and costs $100 (at $24 per hour). Once training is complete, you can open a local server and ask the model any question, from poetry to physics questions like “Why is the sky blue?”
The project generates a detailed report (report.md) with training parameters and comparative results across popular benchmarks, including ARC, GSM8K, MMLU, and HumanEval. While still “kindergarten” performance compared to industrial LLMs, nanochat demonstrates the full functional cycle of a modern model, including interface, evaluation, and user experience.
Karpathy emphasizes that larger versions are under development, priced at $300 and $1,000, which will bring the algorithm closer to GPT-2 levels. The code is optimized for simplicity and transparency: no complex configurations, model factories, or hundreds of parameters. Everything is built around a single, cohesive code base that’s easy to read, modify, and run.
NanoChat can also run on a single graphics card, although it is eight times slower than an 8×H100. For limited GPUs, simply reduce the batch size to avoid running out of memory . The project is entirely based on PyTorch and should work on most supported platforms.
The researcher emphasizes in the paper that nanochat is not just a demo, but a basic, accessible, and reproducible benchmark for studying the architecture of large language models. Its minimalist, open-source design makes it suitable for both students and researchers who want to understand the structure of modern ChatGPT “in miniature.”
Redazione