Can you train an AI model on a home PC?
Yes, if the model is small. On 20 October a developer described on his blog, Ken's Cat Log, how he trained his own copy of GPT-2 from scratch. He did it on his PC at home, with a single RTX 4090 graphics card. The model has 124 million parameters, the settings it learns during training. Training took about 21 hours. By his estimate, the electricity for the graphics card alone cost about 300 yen.
He is not the only one. This summer a 14-year-old developer published Plasma 1.1, a model with 521 million parameters. He trained it from scratch in about nine days on a single RTX 5080, with no cloud computing. He wrote about it on Reddit and on the project's GitHub page.
The recipe itself is public. The Ken's Cat Log experiment follows the code of the AI researcher Andrej Karpathy. In his build-nanogpt project, Karpathy wrote that reproducing the small GPT-2 takes about an hour and about $10 on rented cloud GPUs. What is new is that the job now fits on a desk.
How does it work, in practice?
The post splits the job into three steps. An AI coding assistant, Claude Code, helped him write the scripts and watch the run.
- Public data. He used FineWeb-Edu, a free collection of educational web pages in English, published on Hugging Face under an open licence. Download and preparation took about 12 minutes.
- Base training. The model read about 8.4 billion tokens, small pieces of words, in about 21 hours, by the author's count. At the end it scored 30.5% on HellaSwag, a common-sense test, against 29.4% for the original GPT-2.
- Teaching it to chat. Another 1 hour and 50 minutes on about 460,000 sample conversations turned a machine that continues text into an assistant that answers and then stops.
What doesn't the experiment prove?
It does not prove you can build your own ChatGPT. The author lists the limits himself. The model would claim that Paris has 2.4 billion inhabitants or that 12 plus 7 makes 30. Since it had seen only English, it treated a question in Japanese as text to proofread. His conclusion is that grammar comes early and facts don't. The chat training gave it a conversational tone in five minutes but barely changed how accurate it was.
Why does it matter for your business or public office?
Our read: training an AI model used to mean a data centre. Now a small model can be trained or adapted on one machine in an office. The documents stay in the building, and nothing depends on a cloud provider's prices or terms. Electricity is not the expensive part. The graphics card, the computer and the time of the person who runs the job are, and the post does not price those.
For a business there are three different jobs, and they ask for very different effort.
- Training from scratch. The model learns the language itself, from billions of words. It is what the experiment did, and almost no company needs it.
- Adapting, or fine-tuning. You take an open model that already writes well and teach it your format and vocabulary with your own examples. The experiment's chat step is this kind of job, and it took under two hours.
- Using it with your documents. The model reads your files when it answers, without being retrained. That is where your facts come from, because the experiment shows that adapting changes the tone more than the knowledge.
What about us?
In a typical case we would not train a model from scratch. How we would use it: take a company that answers quote requests every day with the same wording and the same price rules. We would take a small open model, adapt it to the company's past quotes on a machine in its own office, and have it read the current price list when it drafts. A person checks every quote before it goes out. It is the kind of system our AI agents and automation service builds.
So what should you do?
- If you have confidential documents and a repetitive task, such as standard replies or quotes, adapt a small open model and run it on a machine you control.
- If you need broad general knowledge, use a large model through a provider's service (the API) and keep sensitive data out of it.
- If you are not sure, start from your documents, not from training. See what a model gets right by reading them before you adapt anything.
“Train less, read more. Most of what a business needs from AI is already in its own documents, not in a new model.”
DBrothers — how we scope systems
What do people ask about training AI at home?
How much does it cost to train a small model at home?
In this experiment, about 300 yen of electricity for the graphics card, by the author's estimate. The post doesn't count the price of the card and the computer. On rented cloud GPUs, Karpathy put the same small model at about an hour and $10.
Does my data leave the office?
Not if training and use run on your own machine. The experiment trained on one PC at home with public data. With your documents, the same setup keeps them in the building.
Can a small model replace ChatGPT for my team?
No. A model this size writes well but invents facts, as the author shows. It suits narrow, repetitive tasks with a person checking the result, not open questions.
If you want AI working on your documents without sending them out, start with one task and a sample of real files. We will check together whether adapting a small open model on your own hardware is enough.
