Intent detection in 50+ languages
A fine-tuned mmBERT that will run entirely in your browser. No server, no API key — you will be able to open the network tab and check.
Preview — the live in-browser model is being trained. Illustrative output.
—Accuracy vs a large LLM, zero-shot
—Model size after quantisation
—Cost per 1M requests vs an API
StatusBenchmark in progress
How it works
DataAmazon MASSIVE — one million utterances across 51 languages.
ModelmmBERT-small, fine-tuned for intent classification.
RuntimeONNX + transformers.js. Your text never leaves the page.
Notebook, weights and full benchmark will be published on GitHub and Hugging Face.