FUTO Swipe: Open-Source Swipe Typing Models Match Big Tech Accuracy

FUTO has released FUTO Swipe, a family of open-source swipe typing models and algorithms, along with an inference library. The models power the offline FUTO Keyboard for Android, but can be used independently under the FUTO Model License.

Architecture & Benchmarks

FUTO Swipe uses three model types:

• Encoder (635,140 params) – layout-agnostic, language-agnostic, universal predictor.
• ContextLM (1,498,472 params total, 1.1M embeddings) – tiny language model trained per language to filter nonsensical words given context. Only needs text data.
• Decoder (304,155 params) – language- and layout-specific. Currently only QWERTY English trained on real swipe data.

Combined (1,364,271 active / 2,494,767 total params), the models achieve a top-4 fail rate of ~4% on the test set. Excluding out-of-vocabulary words, the error rate is below 1%. This matches big-tech keyboards, according to FUTO. The benchmarks depend on the dataset, but the paper is forthcoming.

Dataset

FUTO released a dataset of 1 million real QWERTY English swipes under the MIT license, collected from voluntary users on swipe.futo.org. Available on HuggingFace.

Inference Library

The accompanying swipe-library (C++, GPL) handles inference, decoding, and dictionary-constrained beam search. With a beam width of 300, it converts swipe paths to ranked word candidates. The library runs on-device in milliseconds, even on low-end devices.

How to Use

• Install FUTO Keyboard v0.1.29 from futo.org for an out-of-box experience with offline swipe typing.
• Download models from HuggingFace and integrate via swipe-library.
• Attribution required per the FUTO Model License.

FUTO is working on a paper detailing training and architecture. The dataset and models are available now for developers to build upon.