Privacy has long taken a back seat in the race for faster text input on phones. Major platforms built swipe typing on vast troves of user data. They send gestures to distant servers. They track habits. They sell the insights. One small organization wants to change that equation.
FUTO released its new swipe typing system this month. The project delivers on-device models that match or approach the accuracy of offerings from Google and Microsoft. Yet it does so without an internet connection. No telemetry. No corporate data hunger. The full details appear on the project page at swipe.futo.tech.
The effort grew from frustration. “For a long time, good mobile swipe typing was locked behind privacy-invasive keyboard apps or unlicensed private libraries,” the FUTO team wrote. They set out to build an open alternative. The result combines three specialized models. An encoder handles general cases across layouts and languages. A small context language model cuts nonsense predictions. A decoder tuned to QWERTY English captures layout quirks and delivers the highest precision.
Numbers tell part of the story. With a beam search width of 300 the system shows roughly 4 percent top-four failure rate on its test set. Error rate falls below 1 percent when out-of-vocabulary words are ignored. Total active parameters sit at 1.36 million. The entire family totals under 2.5 million parameters. Inference runs in milliseconds even on low-end Android devices. Training required nothing more than a single workstation GPU.
Those compact figures matter. They let the keyboard stay truly offline. They reduce battery drain. They open doors for developers who want to embed swipe input in niche applications. FUTO published the models on Hugging Face along with a 1 million swipe dataset collected through a crowdsourced experiment. Contributors visited swipe.futo.org, swiped Wikipedia sentences one word at a time, and consented to public release under an MIT license. The dataset landed in March 2025.
Integration arrived first in FUTO Keyboard, the organization’s fully offline Android input method. An update pushed in recent weeks brought the new swipe engine to users. The app already offered local voice recognition and smart autocorrect. Now swipe joins the stack. Early feedback on Hacker News lit up with enthusiasm. One commenter who had used the keyboard for two years called the improvement “insane.” Others reported 90 to 95 percent accuracy in their own tests.
But praise did not arrive without caveats. Several users noted that word suggestions still trail those from SwiftKey. Random capitalization appears occasionally. Some phrases trigger odd replacements. Developers responded in the thread. One explained that QWERTY creates many colinear letter paths, forcing the model to resolve ambiguity through context and beam search. The team even tested 800,000 synthetic layouts to explore better designs.
Privacy forms the project’s core promise. FUTO Keyboard never phones home. Voice input processes locally. Swipe data stays on the device. The organization sells one-time licenses to fund development and keep the experience ad-free. No subscriptions. No data sales. That stance resonates with a slice of Android users tired of cloud-dependent tools.
The models themselves carry a specific license. FUTO Model Weights License 1.0 grants broad rights to use, modify, and even commercialize derivative work. Attribution must remain visible. The name “FUTO Swipe” cannot be removed from interfaces. The inference library ships under GPL. These terms aim to encourage adoption while protecting the project’s independence. Details sit at the license file on Hugging Face.
Community reaction spread quickly. On X, developer Victor M at Hugging Face highlighted the tiny CNN encoder at 635,000 parameters and 2.65 megabytes that works on any keyboard layout. Posts called the release cool. Some wondered whether the advance signals a larger shift toward open on-device AI for everyday interfaces. Others asked when input might move beyond keyboards entirely.
Discussion on Hacker News, which gathered hundreds of points within hours of posting, revealed both excitement and lingering gripes. One user said the English model felt “quite excellent.” Another admitted ditching the keyboard after months because suggestions felt random. A developer pushed back that the latest decoder had fixed several emoji and capitalization bugs. The thread mixed technical depth with calls for iOS support, better multi-language handling, and custom word memory.
FUTO never claimed perfection. The project page notes that real-world performance varies. The decoder remains English and QWERTY only for now. Future work will expand languages and layouts. A technical paper on training and architecture is forthcoming. In the meantime the C++ inference library stands ready for anyone who wants to experiment. Ideas floated include swipe typing on laptop trackpads or inside virtual reality headsets.
The dataset collection itself broke new ground. Rather than scrape private messages or logs, FUTO asked volunteers to swipe public text. Over one million gestures arrived. Filtering removed low-quality samples. The final release gives researchers and hobbyists a clean starting point. Previous open efforts lacked this scale. Big tech firms guard their swipe corpora. FUTO chose transparency.
Android users can download the updated keyboard from Google Play or sideload the latest APK. Those who buy the license remove any remaining limitations and support continued work. The organization positions itself as a long-term investor in user-controlled computing. Its keyboard forms one piece of that vision alongside other tools.
Whether FUTO Swipe displaces entrenched options remains uncertain. Network effects favor the defaults. Most users never think about their keyboard until it frustrates them. Yet for privacy-conscious developers, tinkerers, and organizations wary of data leaks the project offers a genuine path forward. The models are small enough to run anywhere. The data is open. The code invites extension.
Early signs suggest momentum. Hacker News ranked the announcement near the top. YouTube videos from the FUTO channel explained the release and walked through accuracy gains. Reddit threads in Android communities celebrated the update. One post called it the best improvement the keyboard had seen.
Challenges persist. Suggestion quality must improve to rival years of proprietary tuning. Support for additional languages cannot lag. Edge cases with names, slang, and technical terms need attention. The team knows this. Public development on GitHub and community channels on Zulip and Discord invite contributions.
At root the story is simple. A capability once monopolized by companies that harvest data now sits in the open. The models are fast. They are accurate enough for daily use. They run locally. And they carry no hidden surveillance cost. That combination may prove more disruptive than the raw accuracy numbers suggest.
Developers have already begun to take notice. Some talk of embedding the library in specialized apps. Others explore non-mobile surfaces. If the momentum holds, FUTO Swipe could force bigger players to reconsider their closed approaches. Or at minimum give users a credible alternative that respects their data from the first tap.
WebProNews is an iEntry Publication