tiny_llm in the browser

A small GPT trained from scratch on TinyStories, running entirely client-side. The model and training code live in real-shafe/tiny_llm; this repo is the page that runs it.

Try it: https://tiny-llm.shafe.dev

• The forward pass is plain JavaScript in a Web Worker (public/model.js), with a key/value cache
• GPT-2 tokenizer via js-tiktoken (MIT), bundled into public/vendor/tiktoken-gpt2.js
• Weights ship as fp16. Only the ~8.5K tokens that appear in the training data are kept, which shrinks the download about 5x

Updating the model

With tiny_llm checked out next to this repo and trained (checkpoints/gpt.pt exists):

uv run --project ../tiny_llm python tools/export_weights.py
node tools/check_parity.mjs

The export writes public/model/ and a reference fixture. The parity check runs the JS model against the PyTorch logits.

Run locally

python3 -m http.server -d public 8000

Deploy

Served by a Cloudflare Worker with static assets (wrangler.jsonc). Deploy with:

npx wrangler deploy