Good swipe typing on Android has long been locked inside privacy-invasive keyboards or unlicensed private libraries. FUTO Swipe, released June 13, is an open alternative — a swipe-typing model small enough to run fully offline on cheap phones, and it already ships in FUTO’s keyboard app.
## How it works
The system splits into three tiny models: a universal, layout-agnostic encoder (635k parameters), a language-specific decoder (304k), and a ContextLM small language model (1.5M) for a single language. That’s about 2.5M parameters total — small enough to predict words in milliseconds on low-end hardware, with no cloud round-trip and nothing leaving the device.
## Why it matters
FUTO also released a dataset of 1 million swipes under an MIT license, so others can train and build on it. That’s the real shot here: swipe input has been a quiet moat for Gboard and SwiftKey, propped up by proprietary data nobody else could touch. An open model plus an open dataset makes a genuinely private, offline swipe keyboard buildable by anyone — a reminder that not every useful AI model needs billions of parameters.

Leave a comment