I wanted to try Jev, TypeSafe AI’s new System One model. I got a waitlist.
While I waited, I found Laya. Open weights, a paper, a dataset, an Apache-2.0 licence, and an author who had been publishing this idea since March 2025. It ran on my laptop the same afternoon.
The model was good. The marketing was a Reddit post. So this page is the marketing.
Later I got access to the Jev API and ran both models on the same 500 labelled examples. Jev is more accurate out of the box. Laya is faster from my desk, free, and open, so you can fine-tune it. The numbers are on the main page the way they fell.
If Laya is useful to you, support its author. He has a coffee link on his GitHub profile.