FUTO Swipe ships an on-device neural swipe-typing model

Category: AI
Status: discussion
Sources: swipe.futo.tech, discussion
Summary: FUTO published a neural swipe-typing model used in its Android keyboard and as a standalone library. The system combines a layout- and language-agnostic encoder (635,140 parameters), a layout/language-specific decoder (304,155 parameters), and a small ContextLM (about 1.5M parameters, mostly embeddings) totaling about 2.49M parameters, with beam search (width 300) reporting an approximately 4 percent top-4 failure rate and below 1 percent excluding out-of-vocabulary cases. It was trained on one million voluntarily contributed English QWERTY swipes (August 2024 to March 2025, sourced primarily from Wikipedia), released as an MIT-licensed dataset on Hugging Face. Inference code is GPL; the models use FUTO's own model license.
Comments: HN commenters report accuracy comparable to Gboard and a clear improvement over earlier FUTO versions, with debate over whether the non-commercial model and keyboard licenses meet the OSI open-source definition.
Why it matters: It is a small, fully on-device gesture-typing model with a released dataset and a reusable inference library, an alternative to cloud or proprietary keyboard prediction.
Follow-up: Track multilingual support, the standalone library's adoption, and independent accuracy comparisons.