FUTO Swipe

privacy-first ux only works when the core interaction feels genuinely good, and this is a rare case where it does. keeping the typing path local while still shipping something polished is the right product bet, especially from a team that keeps backing this space.

The split between a layout-agnostic encoder and a layout-specific decoder is the interesting call here, since it implies adding something like AZERTY or a custom split layout only needs a new ~300K decoder rather than retraining the whole stack. Is that actually the case, can I train just the decoder against the released 1M swipe dataset while reusing the encoder, or does the context LM need retuning per layout too? Also curious what the license allows for embedding the C++ runtime into a third-party keyboard.

I've been using swiftkey after trying out a bunch of different swipe-y keyboards a couple years ago. Any plans to launch this on the app store? I'd love to try it out!

Releasing both the models and the 1M-swipe dataset is what stands out here. Beyond keyboards, have you seen developers experimenting with FUTO Swipe for non-traditional input systems like VR, AR, smart TVs, or accessibility-focused interfaces? Curious which unexpected use cases have emerged so far.

shipping open on-device models AND releasing the 1M swipe dataset is the rare open-source move that actually moves the field. respect.

Well done, this is a really worthwhile project, thanks for sharing. There seems to me to be a lot more scope for open-standard solutions being picked up and baked into devices, and maybe even built upon as a store, like Chrome's plugins. That would be a big unlock across devices.

Interesting launch 🚀I like that this runs fully on-device instead of depending on closed keyboard systems or cloud inference.Curious how well the model handles different typing styles over time. Is there any personalization layer planned, where the keyboard can adapt locally to a user's common words, slang, or swipe patterns without sending data off-device?Open models for keyboard input feels like an important direction.

Hi everyone!Swipe typing is one of those features that seems simple, but is actually incredibly hard to get right without relying on closed keyboard systems.FUTO built a three-model system to solve this: a 635K-parameter layout-agnostic encoder, a 300K-parameter QWERTY decoder, and a 1.5M context language model. Because the footprint is tiny, the whole thing can run locally in milliseconds.Releasing the weights and the C++ beam search library means people can now build proper swipe typing for VR, laptops, or alternative mobile platforms without relying on closed-source blobs.The best way to experience it today is FUTO Keyboard.