OPEN LINK ↗
// 1h agoOPENSOURCE RELEASE
FUTO releases open swipe-typing models, library
FUTO has released FUTO Swipe, a family of small, open-weights models and a C++ library for offline, on-device swipe typing. The system runs entirely locally on edge devices and includes a 1 million swipe training dataset.
// ANALYSIS
FUTO Swipe brings high-quality swipe typing out of proprietary black boxes and into the open ecosystem. By optimizing a three-model pipeline to run under 2.5 million parameters, FUTO proves that highly functional edge AI doesn't require massive hardware.
- –Three-part architecture features a layout-agnostic encoder, layout-specific decoder, and a 1.5M parameter ContextLM
- –Achieves a top-4 fail rate of just ~4% on QWERTY English, dropping below 1% when excluding out-of-vocabulary words
- –Models are released under the FUTO Model License, while the C++ inference library is GPL and the 1M swipe dataset is MIT
- –Extremely low resource footprint allows inference on low-end mobile devices and VR headsets in milliseconds
// TAGS
futo-swipefuto-keyboardedge-aiopen-weightssmall-llmopen-sourcedatasetllm
DISCOVERED
1h ago
2026-06-24
PUBLISHED
7h ago
2026-06-24
RELEVANCE
7/ 10
