FUTO Swipe: How an Open-Source Keyboard Just Beat Google Gboard at Its Own Game
Published June 24, 2026 — TekMag
When you swipe across your phone keyboard to type, you probably assume Google’s Gboard — with its billions in R&D, unlimited cloud compute, and decade of data — delivers the best possible accuracy. That assumption just got shattered.
FUTO Swipe, the neural swipe engine inside the open-source FUTO Keyboard for Android, now achieves a 7.38% top-1 error rate in head-to-head benchmarks — a 26% improvement over Gboard’s 11.05%. And it does it with only 1.3 million active parameters, running entirely offline on your device.
This isn’t a hype cycle claim. The benchmarks are published, the model weights are on Hugging Face, and the source code is on GitHub. Here’s how an independent open-source project just out-Googled Google at one of the most fundamental mobile experiences.
The Benchmark That Changed Everything
FUTO’s published benchmark compares four keyboards across thousands of swipe inputs:
The improvement from the previous FUTO version is even more staggering — a 4.6x reduction in error rate. This isn’t an incremental update; it’s a complete rewrite of how the keyboard understands swipes. You can verify the numbers yourself at swipe.futo.tech.
The Secret: Tiny AI, Smart Architecture
What makes FUTO Swipe’s achievement remarkable isn’t just that it beats Gboard — it’s how it beats Gboard. The system uses a three-stage neural architecture, and the entire active pipeline weighs in at just 1,364,271 parameters — roughly 10MB in FP32.
For context, the latest large language models run into the hundreds of billions of parameters. This is the anti-hype AI story, and it echoes other recent breakthroughs we’ve covered, like Moebius: How a 0.22B AI Model Matches 10B+ Giants at Image Inpainting — where a tiny model punches far above its weight class.
The Three-Stage Pipeline
1. Encoder (635K parameters) — A 1D Temporal Convolutional Network (TCN) that takes raw (x, y) touch trajectory data with timestamps and produces per-timestep character emissions. Crucially, this encoder is layout-agnostic: it works with any keyboard layout supplied at runtime using a DCT basis applied to key coordinates. You could switch to Colemak or Dvorak and it still works with zero retraining.
2. Decoder (304K parameters) — A small DFSMN (Deep Feedforward Sequential Memory Network) that refines the encoder’s output for English QWERTY layouts specifically. This is where the labeled swipe data matters most.
3. Context LM (1.5M parameters) — A causal DFSMN with hash embeddings for large vocabularies that handles next-word prediction and beam reranking. Trained on text data alone — no swipe data required.
The entire model was trained on a single workstation GPU using a community-contributed dataset of 1.22 million swipes, collected from volunteers between August 2024 and March 2025 via swipe.futo.org. The dataset is fully open under the MIT license and available on Hugging Face.
This community-driven, efficient model story parallels another open-source AI milestone we covered: GLM-5.2: Open-Source AI Model Beats GPT-5.5 for 1/6 the Cost — proving that open ecosystems can compete with and beat walled-garden AI products.
ClearFlow: A Keyboard Layout Designed for Swiping
FUTO didn’t stop at the neural engine. They also developed ClearFlow — a keyboard layout algorithmically optimized specifically for swipe typing. By testing roughly 800,000 layouts, FUTO engineered a key arrangement that minimizes colinear trigrams (sequences where three keys line up, causing ambiguity in swipe trails). Early adopters report “less frustrating misinterpretations and corrections needed” compared to standard QWERTY keyboards.
The Privacy-First Difference
Perhaps the most dramatic differentiator is what FUTO Swipe doesn’t do. The entire keyboard — swipe engine, autocorrect, next-word prediction, and dictation — runs 100% offline. No data leaves your device. No telemetry. No tracking.
Compare this to Gboard, which sends typing data to Google’s servers for cloud-based predictions, or SwiftKey, which now integrates ads and AI features that require connectivity. FUTO’s approach aligns with the organization’s founding philosophy: “Computers Belong to You.”
This privacy-forward stance is part of a broader trend in AI-assisted software that we explored in The Coming Loop: How AI Harness Engineering Is Quietly Rewriting Software Development — where the next generation of tools puts user agency and local execution first.
The FUTO Ecosystem
FUTO Swipe comes from FUTO (futo.org), a tech-independence organization funded by billionaire philanthropist Eron Wolf. Formerly fronted by right-to-repair advocate Louis Rossmann as Communications Director (he left in February 2025), FUTO also funds Immich (photo backup), Grayjay (creator-focused video platform), Polycentric (decentralized social), FCast, and Live Captions — all with a deliberate anti-VC, anti-surveillance-capitalism mandate.
The FUTO Keyboard app is available on the Google Play Store, via F-Droid, or as a direct APK download from GitHub (latest: v0.1.29.1, June 22, 2026). It uses a one-time purchase model — no subscription — though the licensing is complex: the keyboard app uses FUTO’s Source First License (not OSI-approved open source), while the swipe library is GPLv3 and the dataset is MIT. The $5 billion question: will a licensing model that isn’t fully open source slow community adoption? The HN community certainly had opinions — the discussion peaked at 357 points and #4 on the front page.
Caveats: What FUTO Swipe Still Doesn’t Do
Balanced coverage requires acknowledging where FUTO Swipe falls short:
Android only — No iOS version planned. No timeline.
English QWERTY decoder — The encoder works with any layout, but the refined decoder is English QWERTY only. Other layouts get “good” not “great” accuracy.
No multi-language simultaneous typing — Gboard’s ability to recognize swipes in 2+ languages at once is missing. Users must manually switch languages.
No GIF search — An ideological decision: FUTO considers GIF search a vector for ads and tracking.
App size of 134MB — Mostly voice models (Whisper-based offline dictation). Core keyboard is under 15MB.
UX polish gaps — No swipe-left-to-delete (a SwiftKey staple), occasional random capitalization, and some wonky autocorrect behavior.
The Bottom Line
FUTO Swipe is a genuine David-vs-Goliath moment in mobile technology. An open-source project with 79 contributors, 2,700 GitHub stars, and a community-contributed dataset built one of the most accurate swipe keyboards on the planet — trained on a single GPU, running entirely offline, and beating Google at its own game.
It won’t replace Gboard for everyone — especially multilingual typers and iOS users. But for the millions of Android users who value privacy, want to support independent software, or simply want the best possible swipe accuracy, FUTO Keyboard with Swipe is the new standard.
Available now at keyboard.futo.org.
Featured image: FUTO Keyboard running on Android (source: keyboard.futo.org, promotional material).
Originally published on TekMag.
Top comments (0)