A translation model in a browser tab
the problem
NLLB-200 is not a small model, and running it in the page means competing with the video for memory and compute on the same machine that is decoding 1080p.
what I did
The distilled 600M variant runs through transformers.js on WebGPU where available and falls back to WASM otherwise. Assets are fetched once at install rather than per session, so the cost is paid one time.
