What Seal is, and why the first release is a tool-calling model

Seal will be a tiny, affordable AI device you carry with you. You ask out loud and it gets things done, with no cloud and no account. It can work with your own phone, for things like sending a message, but nothing you say reaches the internet unless you turn on a connection or a tool that sends it.

Seal Research is the part of Seal that builds the models inside it, and this is where I publish them, tested in the open.

Seal is also one person, working alone, with no funding. I would rather say that plainly than let a website imply a team.

What it takes

A device like that does not run on one model. It has to hear you, work out what you want, do it, and answer. Every part has to be small enough to run on the device itself, and good enough that you would trust it. So I build them as separate models, one job each, and test each one in the open, including where it loses. Each runs on Gust, a small runtime I wrote so that a model gives the same output on a laptop, a phone and a browser tab.

The code is MIT and the weights are CC BY 4.0, so you can use, change and ship all of it.

The first release: Relay 1

Relay turns a request and a list of tools into exact function calls, or into no call when nothing fits. It is 11.5 MB. It does not write the call; it reads the request and the tool definitions and points at the answer, so every argument is a value that was in the request.

Across 12 evaluations, run through one harness against 12 other models, its average is 67.2. The best other on-device model scores 48.7. It also loses, on BFCL, ToolACE and NexusRaven, to language models of about 2B parameters and up, and the Relay 1 post has every table, the losses included. The weights are on Hugging Face, the code is on GitHub, and the post has a demo that runs the real model in your tab.

What comes next

Four other models exist and are not released yet. I will publish each one when I can show how it was tested and where it fails.

• Gust Voice, a text-to-speech model of 13 MB that runs about 49 times faster than real time on one CPU core. It is English only, and I would call it a good voice for its size, not a great one.
• Jot, a 16.5 MB speech recognizer that marks the words its own transcript may have wrong. It is still a candidate.
• Mend, 6.5 MB, which takes the ums, repeats and false starts out of dictated text and leaves everything else alone.
• Squint, a QR code reader of half a megabyte that follows a code across frames when the camera shakes.

A second Relay is in research too, aiming to be smaller, faster and more accurate. Whether it gets there is still an open question.

How this is paid for

So far it is me and about €800 of rented GPU time, which I paid myself. If Seal Research’s models are useful to you and you would like me to train the next ones, you can chip in on GitHub Sponsors. It changes nothing about what is released, because the weights are public either way, and a bug report or a benchmark where Relay loses is worth as much to me as money.

If you want to follow the device itself, the waitlist is on the front page.