FUTO releases Swipe, an open swipe-typing model for offline keyboards

Open models for a closed problem

FUTO, the organisation behind the offline Android keyboard FUTO Keyboard, has published Swipe — a family of open models and algorithms for swipe typing. The stated goal is to break a long-standing bottleneck: good swipe typing on phones has been locked inside privacy-invasive keyboard apps or unlicensed private libraries. The models are released under the FUTO Model License, which permits use with attribution, while the inference library is open source.

The system is already shipping: it is available today in FUTO Keyboard version 0.1.29 and later. Developers who want to build their own products can download the models and use the library. A demo on the project site runs server-side to keep the page small; in production, FUTO says, the models run on-device with much lower latency.

Three models and a beam search

The architecture combines three model types. An encoder is layout- and language-agnostic and handles general predictions, but does not reach cutting-edge accuracy on its own. A small, single-language ContextLM removes nonsensical words given the preceding words in a sentence and needs only text for training. A decoder is trained for a specific language and keyboard layout; it learns the layout's peculiarities and reaches leading accuracy. So far FUTO has trained a QWERTY English decoder only.

Predictions alone are not enough: turning a swipe path into words requires a dictionary-constrained beam search that scores candidates. FUTO released swipe-library, a C++ library that handles inference, decoding and beam search so that developers can go from swipe paths to word predictions.

A million swipes, and the numbers

The training data came from the public: in August 2024 FUTO began collecting QWERTY English swipes on swipe.futo.org, where volunteers were given sentences drawn mainly from Wikipedia and asked to swipe them word by word. The effort produced more than a million swipes; after filtering out low-quality entries, FUTO released a dataset of one million swipes under the MIT license in March 2025, and it is available on HuggingFace. FUTO says the data was used both to train the models and to evaluate competing swipe-typing systems.

On FUTO's test set, the full stack with a beam width of 300 produces a top-4 failure rate of about 4 percent; excluding out-of-vocabulary cases, the error rate is below 1 percent. FUTO cautions that such figures depend heavily on the benchmark and that real-world use may vary, but says it believes the system matches big tech's keyboards. The models are small: 635,140 parameters for the encoder, 304,155 for the decoder and 1.5 million for the ContextLM, of which 1.1 million are embeddings — 1,364,271 active parameters out of 2,494,767 in total. A technical report is on arXiv, the models are on HuggingFace and the library is on GitLab.