A 2 MB model that suggests your next word on every keystroke
Chime is the model behind the word suggestions in LocalType, my Android keyboard. Type below and it suggests the next word or finishes the one you’re typing. It runs inside this page, so nothing you type is sent anywhere; the page counts visits and never sees what you type.
Try Chime
Loading Chime, 2.2 MB…
Tap a suggestion to see how many keystrokes it saves.
Chime’s top six, with its probability for each
Probabilities are for the next token over Chime’s whole vocabulary, which includes punctuation and emoji, so they don’t add up to 100%.
Chime in context
Chime has 2.03 million parameters and reads up to 63 tokens of the sentence being typed. It was trained on 161.6 million words of licensed public text, mostly synthetic dialogue, in 2h 2m on one A100. It does not learn from this page.
I trained it from scratch. It is not Gemma or Gemini and is not derived from either, and running it needs no Google model or service: in LocalType it is a 2 MB file read by plain Kotlin, and here by plain JavaScript. The training pipeline is public under Apache-2.0.
The write-up describes its training and measured runtime, and the benchmark how LocalType compares with Gboard at correcting typos and offering the next word. Methods gives the full measurements. Download the weights and NumPy reference.