Ora's own model · runs on this PC · private beta
The small model
The small model
that picks the next click.
Compass-1 looks at the screen, ranks the things it could act on, and picks one. When it isn't sure enough, it hands the call to a bigger model — and anything risky still stops at an approval card.
Tested where it can't cheat.
We measured Compass-1 on websites and apps it had never seen — public pages, no logins — plus two sets held out of its training.
96.8%right on 1,058 decisions across 53 websites and 6 desktop apps it had never seen
100%right whenever it was sure enough to act alone, on that same set
0.07sper decision on a laptop graphics card — about 3s on an older processor without one
Three test sets, drawn to scale.
Where it falls short.
- Renamed buttons cost it about seven points.96.8% dropped to 89.5% when button labels were renamed and shuffled — it leans on what a control's text says. That is its known weak spot.
- Rewording your goal changed nothing.96.8% stayed 96.8% — you don't have to phrase requests a special way.
- It knows when to abstain.On the never-seen set it passed about one decision in four up to a bigger model — and every call it acted on alone was right.
- It is not perfect.About three picks in a hundred landed on the wrong control — which is why anything that can't be undone stops at an approval card, never at a model's guess.
- Slow without a graphics card.About 3 seconds a decision on an older processor. That is why Compass-1 Micro exists.
In the beta
The default. 1.3 GB, downloaded once after you say yes — then it works offline. If the model is missing or can't be verified, Ora falls back to its simpler built-in rules.
In training
A smaller version for laptops without a graphics card, built to answer in about a second on the processor alone. Not in the beta yet.
Built on open models: laya-browser · mmBERT.