An open swipe-typing model under 2.5M parameters, trained on a million volunteer swipes

A swipe model small enough to live on the phone

FUTO Swipe is an attempt to make accurate swipe typing without sending anything to a server. The training data is the unusual part: over a million QWERTY English swipes, collected voluntarily through a website in August 2024, with the sentences drawn mostly from Wikipedia. That dataset is published on HuggingFace under the MIT license — the raw material is open even where the keyboard app isn't.

The architecture is three small pieces working together: a universal Encoder (about 635,000 parameters), a ContextLM that refines guesses using context, and a layout-specific Decoder, with inference done by dictionary-constrained beam search in a C++ library. The total is 2,494,767 parameters — tiny by modern standards. FUTO reports a top-4 fail rate of roughly 4% on its test set, with error rates under 1% once out-of-vocabulary words are excluded.

One reader caught what makes this matter beyond Android: it's, as far as he could tell, the first genuinely useful free-and-open swipe model — which is what would let swipe typing exist on platforms that aren't iOS or Android.

"This is such a massive deal… the first useful free and open Swipe model. This paves the way for things like swipe typing on platforms other than iOS and Android, a major pain point to newcomer OSes." — a commenter who tested it against Gboard

The verdict from people who'd been waiting for this

FUTO's keyboard has existed for years; the swipe quality is what held it back. Several long-term users said this release flipped them.

"I've been using this keyboard on and off for a while now. I've always switched back to gboard, however this update made me convert full time. It's really good." — a user who flagged its remaining quirks

The praise came with specifics, not vibes: people singled out the on-device voice transcription as better than Samsung's or Google's, with automatic punctuation. The complaints were equally concrete — mid-sentence capitalization, weak use of context so an impossible next word still surfaces, and a stubborn habit of swiping "whats" instead of "what's".

Where the demo stumbled, and the license footnote

The browser demo gave detractors ammunition immediately.

"The first two words i tried to type were both 'hello' and I got 'help' and 'hell' as first suggestions." — a new account that tried the web demo

Whether the web demo runs the same model as the shipping keyboard is unclear, which is partly why reactions split so hard between "truly excellent" from multi-day phone users and "worse than HeliBoard" from demo testers. A GrapheneOS user went further, calling the app still alpha-grade and saying he reverted to Gboard with its network access revoked.

The sharpest factual correction was about openness. The "open model" framing is half-true:

"the library uses the GPLv3 (good) while the Android keyboard uses the Futo License (shit)." — a commenter who linked both license files

So the picture is precise: the training data is MIT, the inference library is GPL, and the keyboard you'd actually install carries FUTO's source-available-but-not-free license. For a project whose pitch is openness and privacy, that distinction is the one worth keeping straight.