Why tabi runs on Mistral AI

The story behind tabi, and the five Mistral models that run it.

By Nikhil Bhima, founder of tabi and Mistral AI Ambassador

As a Mistral AI Ambassador, I wanted to show more of what Mistral can do.

Most of what Mistral shows the world is made for companies: workflows, automation, fine-tuning, open models for developers. All useful, and mostly invisible to everyone else. So I kept coming back to one question. What would these models look like in an app people use every day?

Going through the models, I landed on voice. Mistral has models that hear you as you talk, write for you, and speak back, even in a copy of your own voice. Each was good on its own. I hadn’t seen anyone put them all in one app for everyday people.

So I started building tabi, a voice app for the Mac. It was meant to be a fun week of building. I got hooked, and I’ve been making it better every day since.

The name comes from tabby, the striped cat. I wanted a name you could call out to, not a software name, and “hey tabi” just worked.

More than dictation

I didn’t start with a feature list. I read what the models could do first, and built the features from that. Because Mistral could hear, write and speak, tabi became more than a dictation app: it writes for you, rewrites what you select, answers when you ask, and reads to you.

voxtral-mini-2602Hears you, a piece at a time while you are still speakingDictation

mistral-medium-2604Turns speech into writing, answers, rewrites, repliesCleanup, hey tabi

voxtral-mini-tts-2603Speaks text back, in a preset voice or a copy of yoursRead aloud

mistral-ocr-4-1Reads the words off a scanned PDF or a photoListen

DictationHold a key, talk, and the words land where your cursor isVoxtralMedium

CleanupThe ums and false starts come out, and each app can have its own writing styleMedium

Thirteen languagesSpeak one language and write in anotherVoxtralMedium

Hey tabiAsk out loud and the answer lands at your cursorMedium

Voice editSelect text, say what to change, and the rewrite replaces itVoxtralMedium

ReplySay roughly what you mean and the reply gets written for youVoxtralMedium

Read aloudSelect anything and hear it in a real voiceVoxtral TTS

Your own voiceRead aloud in a copy of your own voiceVoxtral TTS

ListenOpen a PDF, a book or a photo and hear it, word by wordOCRVoxtral TTS

Nine languagesHear a document in another languageMediumVoxtral TTS

Why only Mistral

I could have mixed and matched: speech from one company, writing from another, the voice from a third. That’s how most apps like this are built. I chose one.

Trust. tabi keeps as little as it can. The recording is thrown away once it becomes text, your history never leaves your Mac, and there are no accounts. But your voice still has to reach a model to become words. So the real choice was which company to trust with that moment. For me that was Mistral, and I think it is for a lot of people.

Europe. Mistral is a French company, and it keeps data in the EU, under European law. It is one of the few labs of its size outside the United States and China, and plenty of people want a European option for something they talk to every day.

Open. Mistral publishes its speech models, which most labs do not. Its realtime transcription model is Apache 2.0, free for anyone to use. The one that speaks back is published under a non-commercial licence, which is why tabi uses it through Mistral’s paid service.

Price. Mistral’s prices are low enough that dictation in tabi can be free and unlimited, with no account.

Where your voice actually goes

It writes while you are still talking

tabi sends your voice as you say it, and on a long take, every forty or so finished words go off to be tidied while you carry on. By the time you let go of the key, most of your text is already written.

A short sentence lands in about 1.1 seconds. tabi sends your words off to be heard at every pause while you are still speaking, and the tidying happens along the way, so there is little left to do when you let go. A minute of talking lands in about 2.3 seconds. The wait grows far slower than how much you say.

What comes next

tabi is on the Mac for now. Windows is next, then iPhone, then Android. It gets better with every release, and I’ll write about the big changes here.

If you want to try it, tabi is free to download for the Mac.

Questions

Which AI models does tabi use?

Four, all from Mistral. voxtral-mini-2602 turns your speech into text, a piece at a time while you are still talking, mistral-medium-2604 tidies the result and answers “hey tabi” requests, voxtral-mini-tts-2603 reads text aloud, and mistral-ocr-4-1 reads scanned PDFs and photos for Listen.

Why does tabi use Mistral?

Its founder, a Mistral AI Ambassador, built tabi on Mistral’s voice models from the start. It stays on Mistral for trust: tabi keeps nothing itself, and Mistral keeps data in the EU under European law. It publishes speech models openly, its realtime transcription model under Apache 2.0, and its prices let dictation be free and unlimited.

Is tabi made by Mistral AI?

No. tabi is built by Nikhil Bhima, a Mistral AI Ambassador, as his own project. Mistral does not build, own or endorse it. It runs on Mistral’s models through Mistral’s API.

Where does tabi send your voice?

To Mistral, and nowhere else, through a small server tabi runs that forwards the audio and keeps none of it. The recording is thrown away the moment it becomes text, and your history stays on your Mac. The privacy page has the detail.

How fast is tabi's dictation?

About 1.1 seconds for a short sentence, measured from letting go of the key. Because most of the text is written while you are still talking, a minute of talking lands in about 2.3 seconds.

More on the questions page, or see what tabi does.