tl;dr: FUTO has released an open swipe typing system for their offline Android keyboard, trained on a public dataset of 1M+ QWERTY English swipes (released under MIT). The architecture uses three small models—a layout-agnostic encoder, a per-language ContextLM, and a layout-specific decoder—totaling ~2.5M parameters, achieving a ~4% top-4 fail rate (under 1% excluding OOV). Models are under the FUTO Model License (requiring user-visible attribution), with an accompanying GPL C++ inference/beam-search library.