FUTO Swipe - Open models for accurate on-device swipe typing

What it does

FUTO Swipe is a family of small, open models for accurate swipe typing. It includes three model types: a layout-agnostic encoder, a layout-specific decoder, and a lightweight context language model. The full system runs efficiently on-device with a very small footprint. FUTO has also released a dataset of 1 million swipes used to train the models.

Who it is for

FUTO Swipe is primarily developed for FUTO Keyboard, a fully offline Android keyboard app. It is also intended for the broader community of developers who want to build swipe typing into their own applications, such as VR keyboards or laptop trackpad input. The models are released under a permissive license with attribution, and the inference library is open-source.

Why it matters

For a long time, good mobile swipe typing was locked behind privacy-invasive keyboard apps or unlicensed private libraries. FUTO Swipe aims to solve this problem by providing open models and algorithms that can run entirely on-device, preserving user privacy. The models are very small (under 2.5 million total parameters) and can run on low-end devices in milliseconds, with low environmental training costs (requiring only one workstation GPU).

Launch signal

In August 2024, FUTO launched a dataset collection effort on swipe.futo.org to collect QWERTY English swipes. After consenting, users swiped sentences from Wikipedia word-by-word, producing over 1 million swipes. In March 2025, the dataset was released under the MIT license on HuggingFace. The FUTO Swipe models are available under the FUTO Model License, and the inference library is under GPL.

Brand and naming

The name "FUTO Swipe" clearly communicates the product's purpose: swipe typing from FUTO. The brand is positioned as an open, privacy-respecting alternative to proprietary swipe typing solutions. The tagline "Open models for on-device swipe typing" reinforces the key differentiators of openness and local processing.