Skip to content

OpenDecider

OpenDecider OpenDecider

Open, calibrated System 1 decision models. Ask typed questions (choice, score, noul) about any state (text, an email, a ticket or JSON) and get a calibrated probability for every option: 17 ms per question on an NVIDIA GPU, 18 ms on a Mac. Nothing is generated, so there is nothing to parse and nothing to hallucinate. Apache-2.0 code and weights.

pip install opendecider

from opendecider import Choice, Noul, Score, load

model = load("manjunathshiva/opendecider-nano")   # 0.8 GB, downloaded on first use

r = model.system_one(
    "We were billed twice for March. Refund the duplicate today or we cancel our plan.",
    {"department": Choice("Which department should handle this?",
                          {"billing": "invoices, payments, refunds",
                           "technical": "bugs, outages, system errors",
                           "other": "everything else"}),
     "urgency": Score("How urgent is this?", ["not urgent", "soon", "blocking"]),
     "churn_risk": Noul("Does the user threaten to cancel or leave?")})

r["answers"]["department"]["choice"]   # "billing", with a probability for every option
r["answers"]["urgency"]["score"]       # 2 = blocking
r["answers"]["churn_risk"]["noul"]     # probability the answer is yes

Get started Try the live demo Open in Colab

Find a guide

I want to… Read
install it and make my first decision Get started
pick the right model for my hardware and task Choose a model
run a server that existing TypeSafe Jev clients can call Serve it
use LM Studio, Ollama or vLLM as the engine LM Studio, Ollama and vLLM
automate the decisions the model is sure about and send the rest to a person Automate the confident decisions
check an AI agent's next step before it runs Agent guardrails
copy a working script Examples and notebook
see how it compares with Jev, Laya and frontier LLMs Benchmarks
know where it is weak Limitations
see what is coming, or help build it Roadmap
look up a function, an endpoint or a flag Python API · HTTP API · Command line

Why OpenDecider

  • Typed answers, not text. Every answer is a full probability distribution you can threshold, route on or log.
  • Calibrated. Distilled from two openly licensed teachers whose probabilities were temperature-scaled on held-out gold labels. Calibration error on 200 general decisions is 0.083–0.110, against 0.164 for TypeSafe Jev.
  • Runs where you are. CPU, NVIDIA and Apple Silicon from Python; LM Studio, Ollama and vLLM as engines; a server that speaks Jev's /v1/systemone protocol, with Docker images.
  • Measured head to head. Every model, including TypeSafe Jev through its own API, answered the same questions and was scored by the same code. Every number can be rebuilt from the committed results.
typed-decisions (2,000 business decisions) 200 general decisions median latency
opendecider-nano (~400M) 0.796 0.680 17 ms
opendecider-small (4B) 0.672 (zero-shot) 0.735 40 ms
opendecider-medium-td (30B MoE) 0.788 0.765 214 ms
opendecider-large-td (80B MoE) 0.801 0.750 440 ms
TypeSafe Jev 1.13 (API) 0.754 (zero-shot) 0.730 404 ms
Laya's typed-decisions checkpoint 0.766 0.570 21 ms

nano, medium-td, large-td and Laya's checkpoint were fine-tuned on the typed-decisions train split (the test split was never used); Jev and small are zero-shot there. See Benchmarks for the full picture, including where Jev and Laya lead.

GitHub · PyPI · Models on Hugging Face · Live demo · Changelog