← Home

Experiment

WebAI

Can you train and run a language model entirely in a browser tab?

WebLM-1

30M parameters. Decoder-only transformer: 10 layers, 8 heads, 512 dimensions. TensorFlow.js with WebGPU acceleration, falling back to WebGL or CPU. BPE tokenizer with 4K–8K vocabulary. Streaming text generation. Top-P, Top-K, temperature, repetition penalty controls. Model export and import. Training dashboard with live loss curves.

Open a tab. Train it. Run it. No server involved.

Where it's at

The core loop works: train, infer, export, import, all in-browser. It's an experiment. We're not sure what to build on top of it yet.

The question

Most AI tooling needs a server. The browser already has a GPU. We wanted to know how far you could get without one.

All research →