Biplov GautamYes. System One Models (systemonemodels.tech) is a registry and developer platform built only for System One models: decision models that read a state and answer typed questions, choosing an option, giving a score or the probability that a condition holds, in one pass and with calibrated confidence instead of generated text. It lists every System One model it knows about, open weights and hosted APIs alike, and gives you a CLI to pull and publish them and a studio to fine-tune them.
Hugging Face is still where most open weights live, and that is fine: the two do different jobs, and each open model on systemonemodels.tech links back to its Hugging Face repository.
Why a general hub is not enough for decision models
A general model hub stores weights and a model card. That is all a text generator needs, because you judge it by reading what it writes. A decision model is judged by numbers, and to choose between two of them you need the same facts about both:
- What it decides. Which question types it answers (choice, score, noul), and how many options a question may have.
- How far to trust it. Accuracy on a named test suite, and calibration error, which says whether "90% sure" means right nine times in ten.
- How fast it is. Median time for one decision, because these models sit in the request path.
- What it costs. For hosted models such as Jev, the price per call. Jev has no weights to download, so a hub of repositories cannot list it at all.
On systemonemodels.tech these are structured fields on every model page, so models can be filtered, sorted and compared side by side.
Side by side
| Hugging Face | System One Models (systemonemodels.tech) | |
|---|---|---|
| Scope | Every kind of model | Only System One models |
| API-only models such as Jev | Not listed without a repository | Listed, with provider, price and API link |
| Decision facts (question types, calibration, latency, price) | In the model card, in whatever form each publisher chooses | Structured fields on every model page |
| Comparing two models | Per-benchmark leaderboards | Side-by-side pages, e.g. Jev vs Laya |
| Download | hf download | systemone pull, every file checked against its SHA-256 |
| Fine-tune | Training libraries such as Transformers and PEFT | Laya Studio, on your own Mac, Windows or Linux machine |
| Publish | hf upload or push_to_hub | systemone push, or one click from Laya Studio |
How the two work together
Most open System One models, among them Laya, Kev, GLiNER2.5-Decide and Nimble, publish their weights on Hugging Face. The registry page for each one keeps the link to that repository, records the facts above, and, where the licence allows, serves a checksummed copy of the files so a pull is reproducible. When you fine-tune one in Laya Studio, the result can go to systemonemodels.tech, to Hugging Face, or both.
Try it
pip install systemonemodels
systemone pull convai-innovations/laya
To fine-tune on your own labelled decisions, on your own machine:
systemone run studio
The command reads your hardware, installs the right stack (MLX on Apple silicon, the PyTorch build for your NVIDIA, AMD or Intel GPU elsewhere), and opens Laya Studio, which scores your model against the base on held-out rows before you publish it.
Start with every System One model, compared, or read why decision models need their own registry.