What changed
FUTO released Swipe, a set of mobile-oriented encoder, decoder and context models plus a C++ inference library for converting touch trajectories into text. The encoder is layout-agnostic, while the optional decoder and language model are English-focused.
Why it matters
Keyboard teams can test a small, inspectable input model locally instead of sending typing trajectories to a remote service, improving latency and privacy control.
What to do
Review the FUTO Model Weights License, compile the library in a test app and measure word error, layout transfer, memory, battery and offline behavior across consented users and accessibility needs.
Caveat
The checkpoint uses a custom weights license rather than Apache or MIT. English/QWERTY support is stronger than other layouts, and public artifacts do not establish production accuracy across devices or motor patterns.