FUTO has released a swipe-typing system designed to run quickly on low-end devices, together with a C++ library that performs inference, decoding and dictionary-constrained search. The project combines three compact machine-learning components and draws on a dataset of one million English QWERTY swipe gestures collected from consenting volunteers.
Data collection began in August 2024 through a mobile website at swipe.futo.org. Participants received information about the project, consented and then entered sentences, primarily drawn from Wikipedia, one word at a time using swipe gestures. After filtering a small number of low-quality samples, FUTO published one million swipes under the MIT licence on Hugging Face in March 2025.
The released system separates tasks that have different language and keyboard dependencies. A universal encoder interprets the shape of a swipe and is intended to work across layouts and languages, although FUTO says it does not deliver the best accuracy by itself. A small ContextLM uses preceding words to reduce implausible predictions for a particular language. A third model learns the specific characteristics of a language and keyboard layout. Because that decoder needs matching swipe data, the current release supports only English on QWERTY.
FUTO reports a top-four failure rate of about 4% on its test set when all three models are used with a beam width of 300. When words absent from the vocabulary are excluded, the reported error rate falls below 1%. Those figures are the developer’s evaluation; the supplied evidence does not describe the test split, comparison systems or independent reproduction.
The encoder contains 635,140 parameters and the decoder adds 304,155. ContextLM has 1.5 million parameters, approximately 1.1 million of which are embeddings. FUTO counts 1,364,271 active parameters and 2,494,767 parameters in total. The organization says the small footprint permits millisecond execution on inexpensive hardware. Training never required more than a single workstation GPU, according to the release.
Raw model predictions are only part of a usable keyboard. Swipe typing must evaluate multiple possible paths and restrict results to plausible words. The accompanying swipe-library handles that work through dictionary-constrained beam search, scoring candidates and returning word predictions rather than exposing model outputs alone.
The release offers developers an open dataset, modular models and an implementation path for on-device typing. Its present limitation is equally clear: leading accuracy depends on a decoder trained for a specific language and layout, and only English QWERTY is available. Extending coverage will require new consented data and evaluation for each target configuration. The public dataset also makes it possible for outside teams to test alternative architectures against the same collection, subject to careful handling of its scope and quality.
