GPT-2 in Wado

GPT-2 (124M) continues a prompt one greedy token at a time, entirely in your browser. Loam compiled its ONNX graph into Wado, and the result runs as a WebAssembly component. The same code runs from the command line as package-loam/example/gpt2-124m.

The weights are the 548 MB model.safetensors of openai-community/gpt2 (MIT License), fetched from Hugging Face when you ask for them. The page needs Chrome or Chromium 137+ and a few GB of free memory.