How is self-hosting JEMM different from calling hosted Jev?

Hosted Jev answers text-only decisions from an API, while JEMM adds screenshot input and runs the MaestroYan/JEMM adapter on your own hardware with no TypeSafe dependency at runtime.

What hardware and commands stand between me and the first call?

One CUDA GPU with at least 64 GB, installed with pip from git, then python -m jemm.serve with an adapter and port. The model card separately covers use without this server.

What question shapes and image inputs does the endpoint accept?

Choice takes 2 to 32 options, noul takes true/false criteria, and score takes a list of levels. Up to 4 base64 PNG, JPEG, or WebP images ride along and are seen by every question.

When should I refuse to act on the top candidate it returns?

Each answer carries probabilities and confidence. The digest instructs treating confidence below the threshold in decision_config.json as undecided rather than accepting the top candidate blindly.

Is JEMM affiliated with TypeSafe AI?

No. The README states the project is not affiliated with TypeSafe AI, and the entry repeats that.

What adapter does it run?

The MaestroYan/JEMM adapter, served over HTTP with a Jev-style POST endpoint. Runtime has no TypeSafe dependency.

CommunityJev Demos & Experiments

JEMM

Q: How many candidates does it pick and score?

It picks one candidate per question and returns a probability for every candidate in the set.

Open-weight multimodal judgment engine served over HTTP with a Jev-style POST /v1/systemone endpoint, answering choice, noul, and score questions from text or text plus screenshots on a local GPU.

See highlights

Open github.com

Stars: 14
Language: Python
License: APACHE-2.0
Published by: Community
Author: ypcypc
Last verified: Oct 2026

ypcypc/JEMM

User repository on GitHub

View on GitHub

Like Jev, but multimodal and open-weight: picks one candidate per question, with probabilities, from text or screenshots.

Forks: 3
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Like Jev, but multimodal and open-weight. JEMM picks one candidate per question and returns a probability for every candidate, from text or text plus a screenshot, on your own GPU. JEMM stands for Judgment Engine for MultiModal decisions. This repository serves JEMM over HTTP with a Jev-style POST /v1/systemone endpoint; the model card shows usage without this server. Needs one CUDA GPU with at least 64 GB. Not affiliated with TypeSafe AI.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Ranking

Used inJev for Software & DevTools

Languages: Python

Highlights

Picks one candidate per question and returns a probability for every candidate.
Choice takes 2 to 32 options; noul takes true/false criteria; score takes levels.
Up to 4 base64 PNG, JPEG, or WebP images are seen by every question.
Each answer returns probabilities plus confidence with a configured undecided threshold.

Watch out

Needs one CUDA GPU with at least 64 GB and is not affiliated with TypeSafe AI. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

open-jev-typed-decision-engine

Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.