system-one (mpuig)
Open System One model and MLX runtime returning calibrated choice, score and noul probabilities via a Jev-compatible API.
Overview
system-one is an open learning project inspired by TypeSafe's Jev. It fine-tunes small open backbones such as Qwen3-0.6B and MiniCPM 2B to answer narrow, request-defined questions about state with Choice, Score and Noul probabilities, without generating text. Apple MLX is the primary training and serving backend, and the served API follows Jev's /v1/systemone request and answer shape, with a TypeScript SDK.
It is aimed at developers who want a local, inspectable System One model and at people interested in how such models are trained and calibrated. Two tiers are published on Hugging Face: a 0.6B volume tier and a quality-tier MiniCPM 2B q8 build, each with its own temperature file for calibration. The README compares the fine-tuned models with the bare backbone and a pinned Jev version, keeps an audited experiment log with confidence intervals and retractions, and states that some training targets include outputs from a pinned Jev API version. It is not affiliated with TypeSafe.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does the customer explicitly request a refund? | — |
| Choice | What is the main request? | refund · rebooking · ... |
| Score | Request-defined ordered rubric questions over the state(inferred) | — |
How it works
- Serve a fine-tuned adapter with its temperature file on a local port.
- POST state and typed questions to /v1/systemone.
- Receive calibrated probabilities and confidence per question in one forward pass.
- Application code applies thresholds and owns the workflow logic.
Strengths
- Runs locally on Apple Silicon with open weights.
- Jev-compatible API so existing Jev clients can point at it.
- Explicit temperature calibration and a public experiment log.
Caveats
- Requires Apple Silicon for the MLX backend.
- Some training targets come from a pinned Jev version, which is disclosed.
- Learning project, not affiliated with TypeSafe.
Getting started
git clone https://github.com/mpuig/system-one && cd system-one && uv sync, then run scripts/serve.py with the Qwen3-0.6B adapter and temperature file.At a glance
- Type
- Service
- Status
- research
- Language
- Python
- Runtime
- MLX local server
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
An open-source System One decision model — Kahneman.
wh000wh000/awesome-jev-liveAn open-source System One decision model — Kahneman's term for the fast, automatic judgment faculty. Calibrated choice/score/noul probabilities in one forward pass on small fine-tuned open models. Local on Apple Silicon (MLX), Jev-compatible API, audited experiment log.
hellogumbo/awesome-jev
Frequently asked questions
What does system-one decide?
Whatever typed Choice, Score or Noul questions the request defines about the supplied state, returning probabilities rather than text.
How is it different from Jev?
It is an open, locally run reproduction built on small open backbones; Jev is TypeSafe's hosted model.
Does it need a TypeSafe key?
No, it serves its own models locally; the API only mirrors Jev's request shape.
Related projects
- cbjev★ 4
Self-hosted encoder model giving Jev-compatible choice, score and noul answers in one pass, pitched as a better-calibrated Laya successor.
ProjectresearchPython 11 MB CPU-only model that answers choice, noul and score questions with fitted calibration and no LLM at inference.
ProjectresearchPython- kev★ 9,079
Family of small Jev-like decision models you can train and run yourself, with a server compatible with TypeSafe's SDK.
ServicebetaPython Apache-2.0 local decision model (0.8B Qwen fine-tune on llama.cpp) behind the Jev request shape, calibrated and CPU-only.
ServicebetaPythonOpen 596M and 4B decision models that answer Choice, Score and Noul in one pass locally, with calibrated confidence for escalation.
ServicebetaPythonApache-2.0 118M multilingual decision model for Choice, Score and Noul on the System One API, with conformal abstain.
ModelbetaPython