systemonemodels

The command line and Python client for System One, the registry for decision models — models that return a typed decision with a calibrated confidence instead of generated text: choose, score, rank, classify, route.

pip install systemonemodels

This installs the systemone command and the systemone Python package. Python 3.10 or newer.

Sign in

systemone login

The CLI shows a one-time code and opens the registry in your browser, where you approve this machine. Over SSH, open the printed address on any device instead. For CI, skip the browser:

export SYSTEMONE_TOKEN=s1_pat_...   # create one at systemonemodels.tech/settings/tokens
systemone whoami

Find and pull models

systemone search routing
systemone search --capability route --architecture laya --sort downloads
systemone show biplov/snake-balanced-multilingual
systemone pull biplov/snake-balanced-multilingual --variant onnx-int8

Files are cached by content: a second pull downloads nothing, and variants that share a tokenizer store it once. --dest ./snake links the files into a folder of your choosing; systemone cache info and systemone cache clear manage the cache.

Publish

cd ~/my-workspace
systemone push

Run push in a folder and it finds every model underneath, however deep — trained checkpoints, ONNX and Core ML exports, GGUF files, any folder holding model.onnx, model.safetensors, a .gguf or an .mlpackage — lists them, and asks which to publish (1,3-5 or all). Datasets, virtual environments and caches are never searched.

 #  MODEL                        VARIANT        SIZE  FOLDER
 1  banking77-balanced           mlx        646.8 MB  runs/banking77-balanced-0922-234657/model
 2  snake-balanced-multilingual  mlx        646.8 MB  runs/snake-balanced-multilingual-0923-005225/model
 3  snake-balanced-multilingual  onnx-int8  342.7 MB  exports/snake-balanced-multilingual-0923-005225-onnx-int8

Variants of one model are published together, as one version of one repository with a folder each, so systemone pull you/snake-balanced-multilingual --variant onnx-int8 fetches only that folder. A model on its own keeps its files at the top of the repository, the way it sits on disk.

To publish one folder under a name of your choosing:

systemone push ./runs/snake/model --repo you/laya-snake

What gets published, without retyping any of it:

• The model card. The folder's README.md, with its Hugging Face front matter read for the licence, tags and base model rather than shown. A model without one gets a card written from its evaluation, questions and variants.
• The manifest. systemone.yaml — what the model decides, its runtime, and its evaluation: accuracy, calibration error and latency from a Laya Studio run's eval.json, or an export's own measurements. A base model that came from Hugging Face is recorded as such, so the model page links to it. Pass --manifest to use your own, and check it first with systemone validate systemone.yaml.
• The version. The next minor version after the latest (0.1.0, 0.2.0, …), or --version.

push shows the plan before sending anything and, in a terminal, asks first — unless you named the repository with --repo or passed --yes. It warns when a file would publish the path of your home folder. --dry-run prints the manifests, cards and files and stops. --namespace publishes under an organization, --license overrides the card's licence, and --all --yes publishes everything found without a question, for scripts and CI.

Files the registry already holds are never sent again, so republishing with one changed file sends one file, and large files stream from disk with a progress bar that moves by the byte.

What this machine can train on

systemone system          # OS, CPU, memory, GPUs and their driver stack, free disk
systemone system --json   # the same, for scripts

Fine-tune locally with Laya Studio

systemone run studio

Fetches Laya Studio the first time, updates it on later runs, gives it its own Python environment (through uv when it is installed, otherwise a virtual environment on Python 3.11+) and opens it in your browser. It offers to sign you in first, so the studio's Publish button can push a finished run to the registry. Your datasets and runs stay in ~/.layastudio/workspace; the code lives in the platform's data folder and can be deleted at any time.

--port, --no-browser, --no-update and --workspace do what they say, and anything after -- goes to the studio: systemone run studio -- --model aac6fef/laya-mlx. It runs on Apple silicon (MLX), and on Windows and Linux with the PyTorch build for the machine's GPU — NVIDIA (CUDA 13 or 12.6), AMD (ROCm on Linux; RX 7000 and 9000 on Windows), Intel Arc (XPU) — or the CPU. systemone system shows what it found. --source DIR runs a Laya Studio checkout instead of the managed copy.

In Python

from systemone import snapshot_download

path = snapshot_download("biplov/snake-balanced-multilingual", variant="onnx-int8")

path is a local directory with the version's files, served from the cache on every call after the first.

from systemone import Client

with Client() as registry:
    results = registry.search("routing", capability="route")
    model = registry.model("biplov/snake-balanced-multilingual")

Configuration

Development

git clone https://github.com/systemonemodels/systemonemodels-sdk
cd systemonemodels-sdk
uv sync
uv run pytest
uv run ruff check . && uv run mypy

Issues and pull requests are welcome. For anything security-related, see SECURITY.md instead of opening a public issue.

Links

• Registry: https://systemonemodels.tech/models
• CLI guide: https://systemonemodels.tech/docs/sdk
• Manifest specification: https://systemonemodels.tech/docs/manifest
• Changelog: CHANGELOG.md

Licensed under Apache-2.0.