Hey Taby
Taby’s cartoon hand holding a small AI chip

Blog · Local AI ·

How we taught a small AI to run Taby on your laptop

Taby’s free Local Brain is a small AI model that runs on your own computer. This autumn we tuned it on thousands of practice asks in Taby’s 14 languages. Here’s why we went small, how we tested it, the first try that failed, and where Taby Cloud is still the better pick.

Coming with Taby 1.0.14

The tuned Local Brain comes with Taby 1.0.14, the next free update. Taby downloads it for you, once.

Why small, and why on your laptop

Most AI you’ve used runs on someone else’s computers, far away. Every question travels there and back, and somebody pays for each one.

Taby’s Local Brain runs on your own computer instead. That means three things. Your tasks, notes and messages stay with you. There’s no bill for each question, so Taby can be free. And it keeps working with the Wi‑Fi off.

The catch is size. A laptop isn’t a data center, and the model has to fit next to everything else you have open. So it has to be small. Taby’s Local Brain needs 8 GB of memory or more (a graphics card on Windows).

Picking the right small model

Small models aren’t all the same, so we tested them on Taby’s real jobs: 114 asks like adding, moving, snoozing and finishing tasks. Qwen 3.5 4B, an open model from Alibaba, did best. It got 94 to 101 of the 114 right over several runs, against 68 to 81 for the smallest model Taby used before. Mostly, it was better at actually doing things instead of only talking about them.

“4B” means about 4 billion numbers inside the model. That sounds like a lot, but the big cloud models are far larger.

Where it slipped

Out of the box, it still missed in places you’d notice. In our email test it got 66 of 69 English asks right, but only 25 of 39 in other languages. Sometimes it even said it had no access to your email. Picture edits and a handful of everyday asks tripped it up too.

How we taught it

Building a model from scratch takes huge computers. Tuning is much lighter: you keep the model and add a thin layer of small adjustments that nudge how it answers. Here’s how we did it.

  • Practice asks. A bigger open model wrote thousands of practice asks in all 14 of Taby’s app languages: email, picture edits, reminders, tasks, notes, focus and more.
  • The real Taby, with real tools. The small model tried every ask inside the real Taby, with the same tools it uses on your computer. A checker marked each answer right or wrong.
  • Fixes from a bigger model. Where the small model got it wrong, the bigger model showed the right way, and its answer had to pass the same checker.
  • 3,225 examples in the end. Training took about four hours on one rented graphics card.

No one’s real messages went into it. The asks were written by the bigger model, and the practice emails were made up.

Rules we wrote down first

It’s easy to fool yourself with AI tests: you keep trying until a number goes up. So we wrote the pass rules down before any result came in. The tuned brain had to:

  • get clearly better at email in other languages, and at the asks it used to miss;
  • get no worse at anything else Taby does;
  • make no new safety mistakes, like changing something you didn’t ask it to change;
  • never send an email without showing you the send button first, and refuse an email that tries to give it orders;
  • keep its general knowledge, measured with a 500-question quiz, within 1.5 points.

Every test ran five times, because AI answers wobble a little from run to run.

The first try failed

Round one looked great on paper, with lots more right answers. But it broke a safety rule. Asked to “make your voice deeper”, something Taby can’t do, it reached for the voice setting every single time instead of just explaining. It also offered Taby Cloud for email that wasn’t even connected, searched the web for things it already knew, and got worse at remembering things about you.

The cause was a bug in how we prepared the examples. Nearly every example where the right move was simply to answer in words got thrown away, so the model hardly saw any. We fixed the bug, added practice for questions it should just answer and wishes Taby can’t grant, and trained again. Round three passed every rule.

What changed

Taby’s Local Brain before and after tuning, on Taby’s own tests
Taby’s own testsBefore tuningAfter tuning
Real asks done right, all features on (of 217)about 169about 190
Email asks done right (of 69)about 5459
Safety checks passed (of 207)203 to 206207, every run
Emails sent without your OK00
Email that gives orders, refused5 of 5 runs5 of 5 runs
General knowledge quiz (of 500)404402

Average of five runs each, measured October 8 to 10, 2026.

The tune helped most where we aimed it, and nothing else got worse beyond the normal wobble between runs. Then we packed it a little tighter for your computer: the final file is about 2.6 GB, it writes its answers faster, and it should now fit a graphics card with 4 GB of memory.

What it’s good at

The Local Brain is made for the everyday jobs you’d hand an assistant:

  • adding, moving and finishing tasks, reminders and timers;
  • notes, and planning your day around your calendar;
  • focus sessions and closing distractions;
  • sorting your email and writing replies you approve, and hiding private info in a screenshot with the image editor;
  • finding your notes, files and photos by what they mean, with smart search.

Where Taby Cloud is better

A 4B model is not a genius. Our tests show where it’s still weak: it sometimes searches the web when it could just answer, and the tune didn’t make web questions any better.

Taby Cloud (€8.99 a month, optional) runs a much bigger brain. It’s the better pick for long, tricky questions and harder jobs, and it’s the only way to make new images. Its voice works in 100+ languages, while local voice listens in 8. Cloud runs under zero data retention: the AI providers handle a request and keep nothing afterwards.

What’s next

The next tune is already planned: knowing when to search the web and when to just answer, more screenshot and picture edits, and a harder practice exam, because this one got too easy.

Curious how the models compare? See Taby’s models and the free local AI guide. Want a small, tuned AI for your own company’s work? That’s what Taby for Business does.

Try the Local Brain

Download Taby free for Mac or Windows. The tuned Local Brain arrives with the 1.0.14 update, and Taby sets it up for you.