No servers. No API key. No cost per token. A real language model downloads once, then generates text entirely on your device — and this page shows you exactly how fast, in tokens per second.
How this works
This page uses Transformers.js to run SmolLM-135M-Instruct (a small open model from Hugging Face) directly in your browser, inside a Web Worker so the page never freezes. Inference uses WebGPU if your browser offers it, and falls back to WASM on your CPU.
Model: SmolLM-135M-Instruct, 4-bit quantized (q4f16) — small enough to live in your RAM
One-time download: ~121 MB, then cached by your browser for every future visit
Privacy: after the download, nothing you type leaves this page — there is no server to send it to
Cost: $0. Your device does the work, so there is no per-token bill
Honest expectations: a tiny 135M model that lives in your RAM — fun and instant, not GPT-class. It will write plausible text, answer simple questions, and cheerfully make things up. For real API models and what they cost, see the LLM cost calculator.
1. Load the model
Nothing downloads until you click. The ~121 MB one-time download is cached by your browser, then the model runs 100% on your device — free, unlimited, no API key, nothing you type leaves the page.
Not loaded yet.
2. Chat with it
Load the model above, then say hello. Try: “Write a haiku about autumn leaves” or “Explain what a hash table is in one sentence”.