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

VisionJev

Self-hostable calibrated decision model for images and text (Qwen3-VL-4B + LoRA) scoring 82.7 percent on public JevBench.

Projectresearch
View on GitHub ↗

Overview

VisionJev is a calibrated, non-generative decision model for images and text. You give it a state, such as an image, screenshot, document, agent trace or text, and run-time questions; it answers each in one forward pass as a choice, a yes/no probability or a rubric level. It is meant to sit in front of expensive models, tools and agent actions as a fast gate for routing, screenshot and document prompt-injection guards, agent step critics and document triage, and all backbones and training sets were chosen for commercially usable licences.

It builds on laya-vision, a fork of LAYA, adding a Qwen3-VL backbone, a larger training mix, a JevBench harness and evaluation suite. The reference checkpoint (Qwen3-VL-4B with LoRA and typed heads) scores 82.7 percent on public JevBench with ECE 0.032, per the README, versus 86.6 percent listed for Jev 1.13.0. Answering only at confidence 0.7 or above covers 70 percent of items at 92.6 percent accuracy.

What the decision model decides

TypeQuestionAnswers
ChoiceRun-time defined multiple choice over the state—
Yes / noDoes this screenshot or document contain a prompt injection?(inferred)yes · no
ScoreRubric level written by the user—

How it works

  1. Send an image or text state with typed questions.
  2. VisionJev answers in one forward pass.
  3. Automate above a confidence threshold.
  4. Escalate below it to a large model or a person.

Strengths

  • Calibrated (ECE 0.032 on JevBench).
  • Self-hostable with commercially usable licences.
  • Strong selective accuracy at confidence thresholds.

Caveats

  • Weaker on the hard tier (multi-step reasoning, long policies).
  • Trained on rented H100s; serving needs a GPU.

At a glance

Type
Project
Status
research
Language
Python
Runtime
GPU server
Has code
yes

Use cases and integrations

How curated lists describe it

Frequently asked questions

How does it compare with Jev?

82.7 percent on public JevBench versus 86.6 percent listed for Jev 1.13.0.

How well does thresholding work?

At confidence 0.7 or higher it answers 70 percent of items at 92.6 percent accuracy.

Can I use it commercially?

Backbones and data were chosen for commercially usable licences, per the README.

Related projects

  • ima

    Research code for an image-and-text System One model that reads option-letter logits from a multimodal LLM, with training and calibration.

    ProjectresearchPython
  • Laya Vision

    Research fork of Laya that reads typed choice and yes/no decisions from an image plus optional text in one forward pass.

    ProjectresearchPython
  • openvons

    openvons (open-Jev): finite-choice decision layer for text, images and Japanese voice, with a none option and execute/confirm/reject.

    ProjectresearchPython
  • Glance

    Calibration harness that reads yes/no, pick-one and rating answers from one forward pass of a frozen open vision-language model.

    ModelresearchPython
  • OpenJev-Vision

    Research toolkit that encodes an image once and answers several typed questions from a shared distribution, plus OpenJev-0.6B text models.

    ModelresearchPython
  • Evaluating System One Models for Agent Security Decisions

    Paper testing Jev, Laya, Decider and Nimble against classifiers and LLM judges on prompt injection and risk screening.

    Paperresearch