# I built a registry for System One models — here's what I learned comparing all of them

> Source: <https://dev.to/biplov_gautam/i-built-a-registry-for-system-one-models-heres-what-i-learned-comparing-all-of-them-25nh>
> Published: 2026-09-27 15:00:49+00:00

*Disclosure: I'm the founder of systemonemodels.tech, the registry this post is about.*

Ten days after TypeSafe AI shipped Jev and named the category, there were more than fifteen System One model families. A System One model reads a state and answers typed questions — choose an option, score on a scale, or give the probability that a condition holds — in one forward pass, with calibrated confidence and no generated text.

They were scattered across Hugging Face repos, blog posts and API docs, each reporting accuracy on its own test set. So I built one place for them: [systemonemodels.tech](https://systemonemodels.tech/) lists every one with its maker, size, context, licence, price and reported numbers.

Three things stood out:

**The numbers don't rank anything yet**. Every publisher reports on a different suite. The only accuracy that means something is the one you measure on your own labelled examples.

**Small encoders are competitive for narrow decisions**. Laya is a 421M-parameter ModernBERT encoder with a decision head, and it answers a question in tens of milliseconds on a GPU. Fine-tuned on your own labels, it can beat a general model on your decision.

**Lineage matters**. Many models are fine-tunes of fine-tunes. Each registry page shows the full model tree, back to the pretrained backbone.

If you want to try one:

`pip install systemonemodels`

systemone pull convai-innovations/laya

And to fine-tune on your own data, locally:

`systemone run studio`

Everything is free to browse and publish. Feedback welcome — especially on what a fair benchmark across these models should look like.