We have completely overhauled swipe typing to introduce FUTO Swipe, our new swipe typing system that achieves leading accuracy.
This has been a long project in the making, starting with our dataset effort back in late 2024, where thanks to your contributions we built a dataset of 1 million swipes on QWERTY English. It’s thanks to this dataset that we’ve been able to make absolute accuracy evaluations, do various experiments, and develop the system we are releasing with this update.
How accurate is it? In my opinion, a good measure is to look at the error rate (1.0 - accuracy), rather than just accuracy, since it becomes difficult to grasp the numbers that are mostly over >90%. So here are the error rates for a few relevant keyboards, tested on our public test set (filtered to first 20k non-single-letter swipes) in-emulator. A lower number is better, because it means fewer errors. Top-1 error is the rate at which the primary word was incorrect, and top-4 error is when neither of the 3 alternatives were correct either.
FUTO Swipe Gboard iOS Heliboard w/ Google library Old
Top-1 error 7.38% 11.05% 10.82% 13.12% ~34%
Top-4 error 4.19% 5.66% 7.14% 7.63% ~27%
That’s right, we are not matching the big keyboards on this benchmark but rather exceeding them, achieving 26% fewer errors relative to Gboard! Hopefully, swipe typing accuracy will no longer be an issue for you on FUTO Keyboard.
The suggestion bar will now show 3 alternatives after you finish swiping. This is so that you can make full use of the top-4 accuracy (1 accepted word + 3 alternatives). If you prefer the old behavior of showing the accepted word in middle, you can change this in settings. There, you can also find options to change the spacebar and backspace gestures, including disabling the spacebar cursor movement.
We will be working on more evaluations and improvements as we work on a paper, and we welcome feedback about specific cases where the new swipe system is lacking.
In the meantime, FUTO Swipe is also released for you to use in your own open-source projects! We recognize that this technology has been locked behind proprietary and closed-source walls for far too long. To learn more, check out https://swipe.futo.tech/
Emoji compatibility
We added Unicode 17 emoji, and reworked the emoji menu a bit. For a while, you may have experienced missing emoji in the emoji menu if you’re on an older Android version - this is because we depended on the system font to provide emoji, and if an emoji was from a newer Unicode version that wasn’t supported on your system, then it wouldn’t be displayed.
On Android, the typical workaround for this is to use an emoji compatibility library called emoji2, but unfortunately the default implementation requires internet access to dynamically download fonts for missing emoji on the system, which is unacceptable for us as an offline keyboard app. Emoji2 provides a bundled library as an alternative, but this would add a shocking 9MB to the app just to support a handful of emojis on some older devices!
It appears that many of the emojis they bundle are unnecessary, so this new update includes a significantly trimmed compatibility font at only 1MB. Now, you should be able to insert from FUTO Keyboard, even if you’re on Android 7. You may still see tofu rendered by apps - unfortunately this is beyond our control and is on application developers to fix.
The “license” for the swipe typing models. That’s what the link references. FUTO already has their own source-available license for software. And seems now they created one for AI models as well: https://huggingface.co/futo-org/futo-swipe/blob/main/LICENSE.md
That’s interesting, because they say in the video that it’s available under MIT license
https://peertube.futo.org/w/gWLaFZoAnrhmCSeTpzX65F?start=56s
Thats not what their repo is licensed as, so I don’t really believe the video.
Interesting indeed. Their dataset is under MIT license. Their resulting models under a special license.
And I have no clue if there’s code available to customize it, train or fine-tune the models. And what license that would entail.
They often do “interesting” stuff with licensing. Pretty much since the beginning of FUTO. I’d also question the choice of wording in the MIT license. Swipe patterns aren’t exactly “software”. And I wonder if their community knew they were releasing their contributions under some license. I can’t find any paragraph detailing it on the current data contribution page.
Just a bit of a shame how they always frame it in their public relations. Of course they talk about the MIT license in their videos. They seem to known that’s good PR. But they casually drop all the rest and don’t talk about the license of what they did. Or if it’s even actually possible to adapt it to other projects, given that requires some additional code and not just the dataset or weights.
The repo isnt MIT
https://github.com/futo-org/android-keyboard/blob/master/LICENSE.md
So they might have licensed one part as MIT, but not everything. Its still their ‘source available’
Yeah, as I said, they love licenses. Their code is FUTO Source First License 1.kb(?). Their models FUTO Model Weights License 1.0. The dataset MIT and seems the training code is proprietary. That’s 4 differently licensed parts that go into one product.