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

Visual-JEV

SemIf (formerly OpenJev): reads typed option probabilities directly from open models like Qwen3.5-4B, including image inputs, on a home GPU.

Modelresearch
View on GitHub ↗

Overview

SemIf, formerly OpenJev, is an independent research project that reproduces Jev's interface pattern with open models: semantic ifs from open models on a 3090 at home. It reads typed option probabilities directly from a model, with no answer sentence, JSON repair or decoding loop. It does not reproduce Jev's undisclosed model or training.

As of its latest changes, direct, serial and shared scoring accept an optional image path and run Qwen3.5-4B as a vision-language model, and MiniCPM5 2B and Qwen3.5 4B were added to a browser WebGPU demo. A native MLX backend supports Apple Silicon. The scorer outputs typed option scores, timing, the exact model revision and a prompt hash, and refuses to overwrite existing output files.

What the decision model decides

TypeQuestionAnswers
ChoiceTyped option choices from a decisions JSONL file (route, retry, does evidence support X?)—
Yes / noDoes the evidence support X?yes · no

How it works

  1. Prepare decisions as JSONL (optionally with images).
  2. Run semif-score with an open model in direct, serial or shared mode.
  3. Read typed option scores with timing and provenance.
  4. Use results in code.

Strengths

  • Runs on a consumer GPU or Apple Silicon.
  • Image input with Qwen3.5-4B.
  • Browser demo without a waitlist.

Caveats

  • Interface reproduction only; not Jev's model.
  • Requires a GPU able to hold a 4B BF16 model.

Getting started

pip install -e '.[test]' then semif-score --mode direct --model Qwen/Qwen3.5-4B --input examples/decisions.jsonl --output results.jsonl

At a glance

Type
Model
Status
research
Language
Python
Runtime
CUDA, MLX, WebGPU browser demo
Has code
yes
Has tests
yes

Use cases and integrations

How curated lists describe it

  • Multimodal models: Jev-style model built on Qwen3.5-4B that takes images directly, without first converting them to text.

    yibie/awesome-jev

Frequently asked questions

Is this Jev?

No, it reproduces Jev's interface pattern with open models.

Can it use images?

Yes, with Qwen3.5-4B as a vision-language model.

Does it run on a Mac?

Yes, via the MLX backend.

Related projects

  • Diffusion Jev

    Independent Jev-style server on DiffusionGemma and SGLang that classifies doodles, flowers and messages with Choice, Score and Noul.

    ServiceresearchPython
  • privatemode-decisions

    Jev-style choice-plus-probabilities from GLM-5.3-Flash in one forward pass, in Edgeless Systems' confidential-computing API.

    ServicebetaPython
  • RSI-Jev

    Open Jev-style text-and-image decision models built by a self-improving AI research loop, with an effort setting that picks depth.

    ServiceresearchPython
  • TypeAR★ 946

    TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.

    ProjectbetaPython
  • 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