FUTO Swipe: an open-source swipe model that doesn't phone home

What happened

FUTO — the Eron Wolf-funded outfit best known for FUTO Keyboard, Grayjay, and Immich — quietly dropped a new swipe typing model at swipe.futo.tech and rode it to 591 points on Hacker News. The pitch is narrow and obvious: a glide-to-type decoder that runs entirely on-device, ships under a source-available license, and is wired into FUTO Keyboard with no cloud round-trips, no account, and no telemetry.

The technical contribution is a neural decoder trained specifically for swipe trajectories rather than the hand-tuned beam search that older open-source attempts (OpenBoard, AnySoftKeyboard, the abandoned AOSP gesture code) relied on. FUTO's writeup frames it as a sequence model that consumes the raw gesture path plus per-key probabilities and emits a ranked list of word candidates, with language model rescoring on top. The model is small enough to run in real time on a mid-range ARM phone — they're explicitly targeting the kind of hardware GrapheneOS users actually carry, not flagships.

The context matters more than the model card. Google bought Swype's successor tech, then SwiftKey went to Microsoft in 2016, then Microsoft killed the Android SwiftKey beta in 2023 and folded the good parts into Gboard-adjacent products. For roughly a decade, if you wanted swipe typing on Android you had exactly two choices: Gboard (sends keystrokes to Google) or SwiftKey (sends keystrokes to Microsoft). Everything else was either a tap-only keyboard or a swipe implementation so bad you'd give up inside a week.

Why it matters

The interesting thing about FUTO Swipe isn't that it exists — it's that nobody else built it. Swipe typing is a closed problem with public papers going back to ShapeWriter in 2004, and the ML stack to train one of these models in 2026 is embarrassingly cheap. The reason the open-source world didn't have one is the same reason it doesn't have a competitive STT model or a usable open OCR: training data. Real swipe traces are sticky, personal, and Google has billions of them from Gboard telemetry that no FOSS project can replicate.

FUTO's workaround, judging by the writeup and the HN thread, is synthetic trace generation — sample words from a corpus, simulate plausible glide paths through the key centers with noise and acceleration curves, train on that. This is the same trick that made Whisper viable for low-resource languages and that the Pile-style projects use for code. It's not as good as real telemetry. It's apparently good enough.

The community reaction on HN split predictably. The privacy crowd treated it as the missing piece for de-Googled Android — several commenters noted they've been stuck on Unexpected Keyboard or Heliboard and would switch immediately. The skeptics pointed out FUTO Keyboard's license is source-available, not OSI-approved open source, and that the company's broader posture (anti-Big-Tech, funded by a billionaire, somewhat ideological) is not everyone's cup of tea. A third camp — the one worth listening to — pointed out that the actual quality bar for swipe typing is brutal: Gboard has had 15 years of A/B tests on a billion users, and "pretty good" decoder accuracy translates to constant correction friction in daily use.

Early hands-on reports in the thread put FUTO's accuracy somewhere between "clearly behind Gboard on long, ambiguous words" and "indistinguishable on common phrases." That's roughly where Whisper was against Google Speech in 2022 — behind on hard cases, fine on the 90%, and improving fast because anyone can now contribute corrections and retrain. The retraining loop is the unlock. Once a credible open swipe model exists, every fork, every language community, every accessibility researcher can iterate on it. That hasn't been possible in this space, ever.

What this means for your stack

If you ship an Android app with a text input surface, nothing changes for you today — IME selection is the user's call and Gboard's market share isn't moving on a HN front page. But three second-order effects are worth tracking.

First, the privacy-default Android distros now have a real keyboard story. GrapheneOS, CalyxOS, /e/OS — every recommendation thread for these has been hamstrung by "and you'll have to give up swipe typing," which is the single biggest UX regression normal users hit. That objection is gone. Expect a measurable bump in de-Googled Android adoption over the next 12 months, which matters if you're building anything that wants to support that audience (Matrix clients, self-hosted sync, indie payment SDKs that need to not depend on Play Services).

Second, this is a template. The same synthetic-data + small-on-device-model pattern works for handwriting recognition, voice dictation in long-tail languages, and on-device translation — all areas where Google and Apple have moats built on telemetry that FOSS projects could never match. FUTO has effectively published a recipe: pick a closed problem with public research, generate synthetic training data, ship a small model, iterate in the open. If you've been waiting for a justification to look at on-device ML for your product, the friction just dropped.

Third, if you maintain an Android keyboard or input-adjacent project, the model weights and the decoder architecture are now a thing you can build on rather than reinvent. Heliboard's maintainers were already discussing integration within hours of the HN post. AnySoftKeyboard is presumably next. The fragmentation in the open Android keyboard space — five projects, none of them complete — has a path to consolidation around a shared decoder.

Looking ahead

The honest read is that FUTO Swipe is version 0.1 of something that needed to exist five years ago, and the next 18 months will decide whether it stays a niche tool for privacy maximalists or becomes the default in the entire non-Google Android ecosystem. The technical risk is low — swipe decoding is not an open research problem — and the distribution risk is real, because keyboards are sticky and Gboard is preinstalled. Watch the Heliboard integration and the per-language model release cadence. If FUTO ships credible non-English models in the next two quarters, this is the inflection point where "de-Googled Android" stops being a hobbyist phrase.