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

TinyJev

Open 596M and 4B decision models that answer Choice, Score and Noul in one pass locally, with calibrated confidence for escalation.

ServicebetaMIT
View on GitHub ↗

Overview

TinyJev is an open, MIT-licensed decision model family that answers typed questions about text on your own machine. You send state such as a ticket, JSON record or log line, plus questions with the answers you accept, and it returns a probability for each option in one forward pass. It never generates text.

There are two models: TinyJev 0.6B (596M parameters) runs in-process on MLX on Apple Silicon or PyTorch elsewhere, and TinyJev 4B v2 runs in Ollama with up to 16k tokens per request. Both work offline and expose a Python API, a local HTTP server and a System One compatible endpoint. The project is built around calibrated confidence so uncertain cases can be escalated to a person. The README shows recorded comparisons against a hosted LLM writing JSON, reporting similar accuracy on its examples at a fraction of the latency, and a demo driving the Jev Grand Prix racing game.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich team should handle this ticket?(inferred)—
Yes / noIs this statement true of the state?yes · no
ScoreWhere does this state sit on an ordered scale (e.g. priority)?—

How it works

  1. Install the package or pull the Ollama model.
  2. Send state and typed questions via Python, local HTTP or /v1/systemone.
  3. Read per-option probabilities and confidence.
  4. Automate above a threshold; escalate to a human below it.

Strengths

  • Runs fully offline on a laptop.
  • Calibrated confidence supports thresholded automation.
  • The README reports 86 ms per question on a base M1 for the 0.6B model.
  • System One compatible endpoint allows drop-in use with existing clients.

Caveats

  • Smaller model is less accurate than a frontier LLM on some tasks; the README shows 10 of 12 vs 12 of 12 in one example.
  • 4B model is slower on modest hardware (about 5 s per racing decision on an M1 Mac mini).

Getting started

pip install 'tinyjev[mlx,demo]'

At a glance

Type
Service
Status
beta
Language
Python
Runtime
MLX, PyTorch, Ollama, local HTTP server
Has code
yes
License
MIT

Use cases and integrations

How curated lists describe it

  • Open alternative: a 596M pointer-head model that answers `Choice`, `Score`, and `Noul` in a single forward pass and returns calibrated confidence meant to be thresholded, so cases it is unsure about escalate to a human instead of being guessed; MLX-first on Apple Silicon, with a System One endpoint and weights on Hugging Face and ModelScope.

    yibie/awesome-jev

  • A tiny jev-like model that answers Choice, Score and Noul questions in one…

    wh000wh000/awesome-jev-live

  • A tiny jev-like model that answers Choice, Score and Noul questions in one forward pass and returns calibrated probabilities. MLX or PyTorch, fully offline, System One compatible.

    hellogumbo/awesome-jev

  • A small decision model for a laptop that answers the three primitives and is built around knowing when to ask a person.

    MrJev/awesome-jev

Frequently asked questions

Do I need a TypeSafe key?

No, TinyJev runs locally and offline.

Is it compatible with Jev clients?

It exposes a System One compatible endpoint, so clients written for /v1/systemone can point at it.

Which model should I use?

The 0.6B model is fast and runs in-process; the 4B v2 in Ollama reads longer inputs up to 16k tokens.

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
  • ej

    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
  • Gutsy

    Apache-2.0 local decision model (0.8B Qwen fine-tune on llama.cpp) behind the Jev request shape, calibrated and CPU-only.

    ServicebetaPython
  • Verdict

    Apache-2.0 118M multilingual decision model for Choice, Score and Noul on the System One API, with conformal abstain.

    ModelbetaPython
  • Jebadiah

    Open Apache-2.0 System One decision models (4B, 9B, 27B on Qwen) with trainer, data builders, evals and every run record.

    ServiceresearchPython