Build and train a GPT-style language model from scratch, using the same core methodology behind frontier models — at micro scale.
Explained so clearly that even a high school student can build a new language model and train it step by step from scratch, while understanding the core concepts behind how it actually works. Finally, you'll type a line and watch your own words transform into verse.
Type a line.
Watch it become verse.
This model read Tagore, Sarojini Naidu, Dickinson and Poe — nothing else, nothing under copyright. It has 2.7 million parameters, a fraction of a fraction of a frontier model. Every character it writes is computed right now, on your machine, from the checkpoint you pick. Nothing is sent anywhere.
What's actually happening
Weights load once
Your browser fetches the checkpoint's numbers as a flat binary file — about 11MB. After that, everything runs locally.
No server, ever
The forward pass — embeddings, attention, softmax, sampling — is plain JavaScript running on your device. There's nothing to log, nothing to send.
Same math, three moments
Baby, halfway and trained are the identical architecture with different numbers — saved at step 0, step 1,500, and step 5,000 of the same training run.
Heads up: the demo downloads an ~11MB checkpoint the first time you use it.