We have been hard at work improving the user experience, and we're pleased to announce support for the FUTO Swipe engine on our platform. It's an open model for swipe typing built around one simple idea, that your keyboard shouldn't need to connect to the internet, it delivers a privacy-focused typing experience without sacrificing usability.
FUTO has been building an open alternative to the proprietary swipe engines that have dominated the mobile ecosystem for years, guided by a philosophy of open models, open inference, and no vendor lock-in. As they put it in their foreword, quality swipe typing has long been limited to "privacy-invasive keyboard apps or unlicensed private libraries." FUTO Swipe aims to change that by making high-quality, privacy-respecting swipe typing openly available.
The implementation is intentionally lightweight, using a three-model architecture consisting of an Encoder, a compact ContextLM, and a language-specific Decoder. Together, the models total just 2.49 million parameters, allowing inference to complete in just a few milliseconds. On their English QWERTY evaluation set, FUTO reports a top-4 failure rate of approximately 4%. The models were trained on more than one million voluntarily contributed swipe samples, with both the models and the dataset released under the MIT License.
In our current implementation, we are only using the encoder portion of it; since ContextLM and the decoder are language and layout-specific, and we are striving to offer a high-quality experience for everyone in our global community. Even then, without the ContextLM & decoder layers, the results have been impressively good. We hope to be able to train context-awareness models for different languages over time and contribute them back to the world, in order to improve swipe typing for everyone.
What We've Added
Building on FUTO's open-source swipe engine, we've extended it in two key areas:
- Multilingual swipe support: The original implementation launched with support for English on a QWERTY layout. We've expanded the open-source stack to support swipe typing across multiple languages rather than a single language. Spanish is the first additional language.
- A personalized learning dictionary: Swipe input now adapts to your writing over time, improving predictions based on your own vocabulary and typing habits, all processed entirely on-device.
Both features are already running on our development devices. We're currently refining the finger path trail, the visual trace that follows your finger as you swipe to ensure it feels as smooth and responsive as the typing experience itself before it ships.
What's Next
We're targeting the FuriOS 14.0.4 release to bring the FUTO Swipe engine to everyone. With fast, multilingual gesture typing, on-device learning, and a privacy-first design, every keystroke stays on your device.
Stay tuned.