HeliBoard vs Lexo
HeliBoard's README contains the strongest sentence in this whole category: it "does not use internet permission, and thus is 100% offline." An app without that permission cannot send your text anywhere. Not promises not to. Cannot. I make Lexo, Lexo declares the internet permission, and no amount of writing on my part beats that sentence. So this page is mostly about when you should take HeliBoard.
I make Lexo, so read the part about where it loses most carefully. Thirty days free, then $9.99 once. Get it on Google Play
The short answer
- You want a privacy claim you do not have to trust anyone for: HeliBoard. Nothing here beats it.
- You want a language model doing your suggestions and offline dictation in the same app, and will accept a permission set instead of a proof: Lexo.
Why "no internet permission" is different from a privacy policy
Most privacy comparisons are arguments about intent. A company says it does not sell your data, and you decide whether to believe it. HeliBoard's claim is not that kind of claim. Android will not let an app open a socket without the permission in its manifest, so the question stops being about trust and becomes about a fact you can check in ten seconds. It is GPLv3, with Apache-2.0 code inherited from the AOSP keyboard lineage, so you can also just read it.
Lexo declares INTERNET and always will. Three things need it: the one-time model download of about 1.2 GB, Google Play billing for the unlock, and anonymous crash reporting that is off unless you turn it on. None of them is your text. But "you can verify this with a packet capture" is a weaker guarantee than "the app is structurally incapable of it," and pretending otherwise would be the exact move this page exists to avoid.
What HeliBoard gives up
Nothing in HeliBoard's README describes a neural suggestion model or built-in dictation. That is not an oversight, it is the consequence of the design. A language model is hundreds of megabytes that have to come from somewhere, and a speech model is another few hundred. An app with no network permission cannot download either one, so it ships what it can carry.
So voice typing means installing a separate voice input app, and the privacy question for dictation becomes a question about that other app instead. Prediction is a dictionary and a frequency model rather than something that has read a lot of sentences.
What Lexo does with the permission it takes
- Suggestions ranked by Qwen2.5-0.5B-Instruct at q8_0, through llama.cpp on the CPU, inside the keyboard process.
- Dictation by an NVIDIA Parakeet transducer, int8 ONNX through sherpa-onnx, on the phone.
- Transcript cleanup by a small ELECTRA-based tagger with three heads plus a rule pass, so fillers, casing and punctuation get fixed without a generative model rewriting what you meant.
- Five permissions total: RECORD_AUDIO, VIBRATE, POST_NOTIFICATIONS, INTERNET, ACCESS_NETWORK_STATE.
All of that inference runs locally. The permission buys the download and the purchase, not a round trip while you type.
The honest framing
These are not really competitors so much as two different answers to "what am I willing to give up." HeliBoard gives up model quality to get a claim that needs no trust. Lexo gives up that claim to run models. If you cannot decide, the tiebreaker is whether you would actually use offline dictation. If you would, you need a keyboard that can download a speech model. If you would not, take the stronger claim.
Two more things that go HeliBoard's way and are worth saying plainly. It is open source and Lexo is not. And it is on F-Droid, so it installs on GrapheneOS and LineageOS, where Play Integrity currently blocks Lexo entirely.
Lexo is the keyboard this is about. Thirty days free, then $9.99 once, no subscription and no ads. Get it on Google Play
More: six Android keyboards on what each actually sends, running a 0.5B language model inside a keyboard, what offline dictation actually means, and is Lexo safe.