Lexo

Thoughts in.
Words out.

An Android keyboard that predicts with a language model running on the phone, and types what you say the moment you say it. All of it offline.

Free for 30 days, everything unlocked. Then $9.99 once, not a subscription.

Built on the open-source FlorisBoard project. Five permissions, and none of them send what you type.

Next word, ranked on the phone

You typed

The quarterly review is scheduled for next Thursday at

It offers

noonthemidnighthalfoneten

You said

um can you uh send me the deck tomorrow morning

It typed

Can you send me the deck tomorrow morning?

Cleanup is optional and ships off by default; the toggle is in Settings › Typing.

Ways in

  • Tap

    A language model reads the context and ranks what you're likely to type next, on top of a word-frequency prior and your own vocabulary.

  • Speak

    Hold the mic and talk. An offline recognizer types what you say as you say it, and an optional cleanup tagger strips fillers and restores casing.

  • Swipe

    Swipe through a word instead of tapping it out. Glide typing, decoded by a neural classifier rather than a dictionary lookup. Part of Pro, so the trial covers it.

Lexo keyboard with a suggestion strip above the keys offering the next word in a half-typed sentence
Suggestions ranked on the phone.
Lexo dictating: the mic key is recording and speech is landing in the text field as a punctuated sentence
Recognition runs locally.
Lexo in dark theme, suggesting the next word of an unfinished message
Dark theme, same strip.

What's actually running

  • Suggestion model

    English runs a fine-tuned SmolLM2-360M with a word-scoring head, about 431 MB; other languages run gemma-3-1b-it, Qwen2.5-0.5B-Instruct or their own word-level pack. All of them through llama.cpp on CPU, inside the keyboard process, one language resident at a time.

  • Speech recognizer

    English, Spanish, Portuguese and French use NVIDIA Parakeet, ~0.6B, int8 ONNX through sherpa-onnx with a silero voice-activity gate, about 661 MB. Parakeet ran at roughly 0.05 real-time factor on a Pixel 9 Pro XL, a ten-second window decoding in about half a second.

  • First-run download

    Setup pulls the typing model for your language, about 431 MB in English, over Wi-Fi, verified by md5. Voice is opt-in and downloads separately the first time you use the mic (about 661 MB of speech model plus a 14 MB tagger in English). The models live in the app's private files directory, so app updates stay small and never re-download them.

The full model list per language, and the latency numbers measured on the 1.0.3 (4199) build, are in the build note and the dictation write-up.

None of it leaves the phone

The models run locally, so your text and your audio have nowhere to go. That's a claim like any other, and it should be checkable rather than promised.

Lexo declares exactly five permissions: RECORD_AUDIO for dictation, VIBRATE for haptics, POST_NOTIFICATIONS for download progress, plus INTERNET and ACCESS_NETWORK_STATE. So point PCAPdroid or NetGuard at Lexo: past the model download, typing and dictating generate no traffic at all.

Longer version, including what I can't prove to you: is Lexo safe, and what does it collect.

Where it's weak

  • Swipe typing is new. It shipped in August 2026, and Gboard and SwiftKey have had a decade to tune theirs.
  • English is much stronger than the rest. Seven other languages ship, and the gap is real. Closing it is the current focus.
  • The download is big. About 431 MB in English to set up typing, and roughly 660 MB more the first time you use the mic, with some other languages larger still.
  • De-Googled ROMs are rough. Licensing and billing verification lean on Play services, so purchases can fail to activate on Huawei, Aurora, LineageOS and GrapheneOS.
  • Dictation cleanup is off by default. It isn't good enough yet, so fresh installs get the verbatim transcript.

If none of that puts you off, the trial is the fastest way to check the rest.

Get it on Google Play

Try everything for 30 days

After the trial, if you don't buy

An ordinary keyboard

  • Layouts, language switching, emoji, clipboard history, themes
  • One-handed mode, and key sounds, vibration and gestures you can set
  • No suggestion strip and no autocorrect: those run on the models, and the models are the paid part
  • No countdown, no locked keyboard, no ads

Being straight about it: Lexo is a paid keyboard. Almost everything that makes it different from the keyboard you already have is behind the unlock, and what an unpaid install keeps is an ordinary keyboard.

Getting started

  1. Install it from Play. The app itself is small.
  2. Enable Lexo, then switch to it. Android puts this under Languages and input, On-screen keyboard.
  3. Join Wi-Fi once and let the typing model download. About 431 MB in English, one time. Voice downloads separately, the first time you use the mic.

Thirty days free, then $9.99 once.

Get it on Google Play

Questions

How do I switch to Lexo after installing it?

Android hides this in two steps. Enable Lexo under System settings, Languages and input, On-screen keyboard, then tap the keyboard icon in the navigation bar to switch to it the first time. Lexo walks you through both on first launch.

Is it a subscription?

No. Everything is unlocked free for 30 days, and after that the unlock is $9.99 once. There's no recurring charge, and no ads at any point.

Does it work without an internet connection?

Yes, once the one-time model download is done. Typing, prediction and dictation all run on the phone, so airplane mode is the simplest way to check that for yourself.

Does anything I type get sent anywhere?

No. Lexo declares five permissions and none of them carry your text or your audio. Point PCAPdroid or NetGuard at the app: past the one-time model download, typing and dictating generate no traffic at all.

Writing

30 days free
then $9.99 once

Get it on Google Play