FUTO Releases Open Swipe Typing Model and Dataset
Original: FUTO Swipe – A new swipe typing model
Why This Matters
Open-source swipe typing addresses mobile keyboard privacy concerns while matching proprietary competitors' accuracy.
FUTO released FUTO Swipe, an open-source swipe typing system with 1 million English swipes dataset under MIT license. Models run fully offline on Android devices with 4% top-4 fail rate and 2.5 million total parameters.
FUTO introduced FUTO Swipe, a family of open models and algorithms designed to provide privacy-preserving mobile swipe typing without relying on proprietary keyboard apps. The system comprises three model types: an encoder model (635,140 parameters) for general layout-agnostic predictions, a ContextLM language model (1.5 million parameters) for single-language context refinement, and a decoder model (304,155 parameters) for layout and language-specific accuracy. Combined with beam search width of 300, FUTO Swipe achieves approximately 4% top-4 fail rate on test sets, with error rates below 1% when excluding out-of-vocabulary cases. FUTO developed a 1 million swipe dataset collected from August 2024 through March 2025 via voluntary contributions on swipe.futo.org, with sentences primarily sourced from Wikipedia. The dataset was released under MIT license on HuggingFace in March 2025. The inference library, swipe-library, written in C++, handles dictionary-constrained beam search and full inference pipeline. Models run entirely on-device with millisecond latency on low-end devices. FUTO Swipe models are licensed under the FUTO Model License with attribution required, while the inference library uses GPL licensing. The company trained all models using only a single workstation GPU, minimizing environmental costs.