Not every model should write. Some should decide — fast, typed, with a confidence your code can branch on. This is where those models, and the people who build them, live.
Built for models that decide. Not adapted from tools for chat.
examples /
The right request. The right destination. A choice model for support ticket routing.
“I can’t connect to the API after rotating my key.”
01 / A DIFFERENT KIND OF MODEL
Language models are the slow, deliberate System 2 of AI: open-ended, and at their best when they write. System One models are the fast half. Given a situation, they return one decision and how sure they are of it. Different outputs, different metrics, different tooling — so a different home.
Which way should the snake go next?
The snake is heading east, and the food looks to be above it and slightly to the left. Turning left would bring it closer, but it is worth checking first whether its own tail could block the way back. On balance, the most reasonable choice is probably to turn left — although carrying straight on for one more step would also be defensible.
{"action": "right",
"confidence": 0.89}Illustration of each model’s output. Latencies vary by model and hardware. Read the full argument
capabilities: choice · score · rank · classify · routedecision_accuracy · calibration_error · valid_action_ratemedian_latency_ms · p95_latency_msbase_model · runtime · variants per folder02 / THE FOUNDATION
A focused home for a new class of models.
Phase one brings the essential pieces of
the decision-model workflow together.
Discover models by decision primitive, architecture, and license. Start with what your application needs to decide.
A model card, a common manifest, and every file versioned. The context you need, before you write the first line.
Trace a fine-tune to its base model. Give your work a stable namespace and make it discoverable by the next developer.
Your files live here, versioned, or where they already are. Push weights straight to the registry, or reference Hugging Face, GitHub and any URL.
03 / A COMMON LANGUAGE
Meet systemone.yaml. A proposed manifest that makes a model’s capabilities, runtime, and lineage easy to understand.
model: support-routernamespace: examplescategory: system-onearchitecture: example-runtimelicense: apache-2.0capabilities:- choice- scorebase_model: examples/baseruntime:framework: pytorchartifact: model.safetensors
04 / FOR THE PEOPLE WHO BUILD THEM
Most model tooling assumes the output is text. If yours is an action, a label or a score, you have been making do with tools built for someone else. This one is built for you — publish yours, pull someone else’s, build on each other’s work.
Policies that pick the next move — like the snake agent already on the registry.
Controllers that choose an action many times a second, on the device itself.
Send the ticket, the request or the call to the right place, with a confidence to escalate on.
A calibrated score instead of a paragraph, so a threshold means what it says.
Order the options and return scores your product can sort by.
Small enough for a phone or a browser, fast enough to run on every event.
One identity and one repository model, from the first pull to the fine-tune you publish back. Sign in through the browser once; after that it is push and pull, with files cached and shared between variants.
$ pip install systemonemodels Successfully installed systemonemodels-0.1.0 $ systemone login Your one-time code: BCDF-GHJK Opened systemonemodels.tech/device in your browser. ✓ Signed in as you $ systemone push ./snake-agent --repo you/snake-agent Published you/[email protected] with 9 files $ systemone pull you/snake-agent 9 files · 343.1 MB from cache
05 / WHERE THINGS STAND
The registry and the CLI are live today.
What comes next depends on what you build with them.
Discover, understand, and publish model repositories.
Log in through the browser, then push and pull from your terminal.
What comes next is shaped by the people publishing here.
We’re building its home.
# Illustrative draft — schema subject to change