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

Vev

Open Jev-style decision models with vision (Qwen3.5 4B/9B LoRA) serving Choice, Score and Noul on /v1/systemone.

LibraryresearchApache-2.0 (code), CC BY-NC 4.0 (weights)
View on GitHub ↗Live demo ↗

Overview

Vev is an open Jev-like decision model that can also see images. It is a LoRA fine-tune of Qwen3.5 at 4B (about 10 GB GPU memory) and 9B (about 19 GB), in English and Chinese. Screenshots and photos go into the state alongside text, and each Choice, Noul or Score question gets a probability for every allowed answer, read from label-token probabilities with nothing generated. The README cites about 40 ms per answer.

It serves the /v1/systemone format, so TypeSafe's SDK works by changing the base URL. The README shows a checkout screenshot example and measurements on reading UI state, checking images against a written policy and checking generated images against prompts. A Doom demo slices the screen into eight vertical strips and asks one yes/no per slice. Code is Apache-2.0, but weights are CC BY-NC 4.0, non-commercial only.

What the decision model decides

TypeQuestionAnswers
Yes / noDoes the screen show an error message?yes · no
ChoiceWhich checkout step is the user on?shipping · payment · review
Yes / noDoes this screen slice show a monster?yes · no

How it works

  1. Run Vev on your GPU.
  2. Send text and images in the state with typed questions.
  3. Read per-option probabilities.
  4. Threshold and act in code.

Strengths

  • Multimodal input with Jev API compatibility.
  • Fast enough for real-time use (about 40 ms).
  • Open code and weights.

Caveats

  • Weights are non-commercial only (CC BY-NC 4.0).
  • Needs 10 to 19 GB of GPU memory.

At a glance

Type
Library
Status
research
Language
Python
Runtime
GPU server (/v1/systemone)
Has code
yes
Has tests
yes
License
Apache-2.0 (code), CC BY-NC 4.0 (weights)

Use cases and integrations

How curated lists describe it

  • Open alternative: an open-source Jev implementation with vision input, LoRA fine-tuned on Qwen3.5-4B and 9B, serving `Choice`, `Score` and `Noul` questions on the `/v1/systemone` wire format with screenshots and photos placed directly in the state so one decision can use both text and image; weights are CC BY-NC 4.0, non-commercial only.

    yibie/awesome-jev

  • Open-source Jev implementation with vision input, fine-tuned from Qwen3.5 4B / 9B. Screenshots and photos go straight into the state, and Choice, Score, and Noul answers draw on both text and image, read from label-token probabilities with no generated text. Serves `/v1/systemone`; open code and weights, weights for non-commercial use only.

    AnotiaWang/awesome-decision-models

  • Open-source Jev implementation with vision input: 4B and 9B LoRA fine-tunes of Qwen3.5 that take screenshots and photos in the state alongside text, and return a probability for every allowed answer of each Choice, Noul or Score question, read from label-token logits rather than generated text. Serves the `/v1/systemone` format, so TypeSafe's SDK works after changing the base URL; conformance runs and evaluation results are in the repo. Code Apache-2.0, weights CC BY-NC 4.0 (non-commercial).

    cobanov/awesome-jev

  • Jev-style decision models that also take images, with open weights on Qwen3.5 for your own GPU.

    MrJev/awesome-jev

Frequently asked questions

Can I use it commercially?

The code is Apache-2.0 but the weights are CC BY-NC 4.0, non-commercial only.

Does it work with TypeSafe's SDK?

Yes, by changing the base URL.

What hardware is needed?

About 10 GB GPU memory for 4B and 19 GB for 9B.

Related projects

  • OneJev

    OneJev: open multimodal System One model (0.8B-27B) answering typed questions about screenshots, photos, video and text in one pass.

    ModelbetaPython
  • amigos-jev

    Free hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.

    Servicebeta
  • ComfyUI-TypedDecision

    ComfyUI nodes for Jev-style noul, choice and score decisions over images and text using a local imajev adapter on Qwen3.5.

    ProjectbetaPython
  • 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
  • jev-spatial

    Independent open model on Molmo2-ER that answers spatial questions by choosing from fixed options with one shared classifier head.

    ProjectresearchPython
  • Jevlet

    From-scratch research reconstruction of a small System-One decision model, used to route and gate commands in a Windows palette.

    ProjectresearchPython