Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

oev

OEV: 184M-parameter decision engine that scores options in one pass, reporting 0.7705 on typed-decisions and 22 ms per question.

ProjectresearchApache-2.0
View on GitHub ↗Live demo ↗

Overview

OEV is a small neural decision engine. Instead of generating text it packs the state, question and every option into one sequence and returns a calibrated probability distribution from a single forward pass. Each option is embedded as its own anchor with full tokens, which the author says lets accuracy scale with label count.

The README reports 0.7705 on the typed-decisions benchmark for a single 184M model, slightly above Laya's published 0.766 from a 421M checkpoint, 0.7760 for a four-checkpoint ensemble, 0.8594 on Banking77 from one file (Jev leads there at 0.870), 0.8650 zero-shot on DAIR Emotion, and 22.2 ms per question on a T4 GPU, with an ONNX INT8 CPU build at 54.2 ms p50. Weights are Apache-2.0 on Hugging Face, with a demo space, a PyPI package and CI tests. It is for builders wanting a compact, self-hosted decision model.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich option fits the state and question?—

How it works

  1. Pack state, question and options into one sequence.
  2. Run one forward pass of the 184M model (GPU or ONNX INT8 CPU).
  3. Read the calibrated distribution over options.
  4. Act on the top option or threshold on confidence.

Strengths

  • Small and fast: 22.2 ms per question on a T4 per the README.
  • CPU deployment via a 228 MB INT8 build.
  • Open weights, tests and CodeQL in CI.

Caveats

  • Benchmarks are author-reported.
  • Trails Jev on Banking77.

Getting started

pip install oev, or download weights from https://huggingface.co/divyanshudhruv/oev-typed.

At a glance

Type
Project
Status
research
Language
Python
Runtime
PyTorch, ONNX
Has code
yes
Has tests
yes
License
Apache-2.0

Use cases and integrations

How curated lists describe it

Frequently asked questions

How big is OEV?

184M parameters, about 0.44x the size of Laya's 421M checkpoint.

Can it run on CPU?

Yes, the ONNX INT8 build runs at 54.2 ms p50 on 8 threads per the README.

How does it compare with Jev?

The README says Jev leads only on Banking77 (0.870 vs 0.8594).

Related projects

  • sokudan

    Japanese System One decision model on ModernBERT-ja (314.6M) returning typed choice, score and bool distributions with no generation.

    ProjectbetaPython
  • GLiNER2.5-Decide

    Fastino's 340M Apache-2.0 decision classifier that scores caller-defined labels in one forward pass for routing, intent and scoring.

    ModelproductionPython
  • DeepOpen

    Open multilingual System 1 decision engine built on Laya checkpoints with a router that picks the best checkpoint per request.

    ModelbetaPython
  • brier

    Mangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.

    ProjectproductionPython
  • blink★ 22

    Open one-pass typed-decision model with a dependency-free C99 runtime that also builds to a 66 KB WebAssembly module.

    ServicebetaC
  • Foq★ 11

    Local open alternative to Jev using a 1.58-bit ternary 8B model to answer booleans, choices, scores and Pydantic extraction in ~25 ms.

    ProjectbetaPython