You talk
Speak whenever you like. From another app, press ⌥ Space or say “OK Maé”. Your words show as you speak.
You talk, Maé answers out loud, in real time. Everything runs on your Mac: your voice never leaves your devices.
In development No download yet.
No button to hold, no set phrases: you talk to her like you would to anyone.
Speak whenever you like. From another app, press ⌥ Space or say “OK Maé”. Your words show as you speak.
The language model runs on your Mac’s chip, not on a server.
She starts talking before her answer is fully written. Want her to stop? Just speak.
Speech recognition, the language model and the voice all run on your Mac. No account, no telemetry.
The details, feature by featureMaé answers your questions, and can do much more if you let her.
Press ⌥ Space, or your own shortcut, and talk. A small capsule follows the conversation without hiding what you’re doing.
Hands-free: say her name, then what you want.
OK Maé, what time is it?
She adds them to your calendar and reminders, and runs your shortcuts.
Set a ten-minute timer for the pasta.
The news, a score, opening hours, a price: she searches with the engine you choose, and tells you where the answer comes from.
What time does the pool close on Sundays?
Precise answers, from dedicated services rather than a search.
Will it rain in Lyon tomorrow?
She remembers what you ask her to remember, and nothing else.
Remember that I’m allergic to nuts.
She answers in the language you speak, and switches on her own.
Quel temps fait-il à Paris ?She hands a website, code or research to an agent like Claude Code, and a notification tells you when it’s done.
Home Assistant or your scripts can send her a sentence to say. It only informs her: it never triggers an action.
“OK Maé”, the actions, web search, memory and the iPhone stay off until you turn them on in Settings.
The iPhone and iPad app becomes your Mac’s microphone, speaker and screen. Talk to Maé from anywhere: the work stays on the Mac.
The iPhone and iPad app isn’t on the App Store yet.
A recent Mac is enough. The more memory it has, the larger the language model.
| Memory | Language model | Download | Tests |
|---|---|---|---|
| 8 GB | Gemma 4 E2B | 5.7 GB | Estimated |
| 16 to 24 GB | Gemma 4 E4B | 7.3 GB | Estimated |
| 32 to 36 GB | Gemma 4 26B-A4B | 17.5 GB | Tested |
| 48 GB and more | Gemma 4 26B-A4B, 8-bit | 30.1 GB | Estimated |
Only the 32 to 36 GB tier could be measured so far, on an M1 Pro. The app picks the model that suits your Mac.
Open models, running on your Mac’s chip. The app is written in Swift.
Voice detection
Silero VAD
On the CPU
Speech recognition
Parakeet TDT v3
NVIDIA · on the Neural Engine
Language model
Gemma 4
Google · on the GPU, with MLX
Voice
Pocket TTS
Kyutai · on the GPU
from the end of your sentence to the first sound of her answer.
for her to stop when you talk over her.
Measured on an M1 Pro Mac with 32 GB. For the answer, a median; the goal is to get under one second.
There’s no date yet. OpenMaé is in development, and there’s no download yet. To try it now, you build the app from its source code: everything is explained on GitHub.
Yes. You need the internet once, to download the models. After that she works offline, except for the features that go online, if you turn them on.
Your Mac, and only your Mac. The microphone’s sound is processed on your Mac and sent nowhere. With the iPhone app, it travels encrypted to your Mac.
French first, and English. She answers in the language you speak to her. The app itself is in French and English.
You need a Mac with Apple silicon (M1 or later), macOS 15 or later and at least 8 GB of memory. Intel Macs are not supported.