How does this compare with calling hosted Jev directly?

Hosted Jev handles text decisions through an API, while this is a self-hosted multimodal option you serve locally. It accepts images and video alongside text and exposes a compatible request shape.

Which serving backend should I pick for video tasks?

Use the PyTorch backend for video and the full modality set on NVIDIA GPUs. The lighter-weight path covers text and images only, so video work needs PyTorch.

Can I fine-tune it on my own decision data?

Yes. The repo publishes a large question set with media, unpack scripts, and per-size configs; point the config at your own training data and media root and serve the resulting checkpoint the same way.

What does local serving look like from client code?

Both backends serve the same local API, and the Python example sends state, media, and questions to localhost. Larger sizes need sharding-aware configs and more GPU resources.

Which OneJev backends handle video, and which do not?

PyTorch on NVIDIA GPUs supports text, images and video, while llama.cpp runs GGUF models for text and images only. Use PyTorch for video. Docker Compose ships torch, rocm, vulkan, gguf, gguf-rocm and cpu profiles, all serving the API and playground on port 8000.

How fast is OneJev on a 1280x720 screenshot?

On one H200 the 0.8B model answered one question in 31 ms and ten in 51 ms, 4B took 64 and 104 ms, 9B 81 and 131 ms, and 27B 189 and 324 ms.

How was OneJev trained, and what data is released?

The four sizes fine-tune Qwen3.5-0.8B, 4B and 9B plus Qwen3.8-27B for one epoch with the vision tower frozen, at learning rate 5e-6, a 16,384-token limit, and cross-entropy plus Brier loss. OneJev-Data releases 94,707 of the original 99,193 training questions.

CommunityJev Demos & Experiments

OneJev

Community multimodal System One model in sizes from 0.8B to 27B that answers typed Noul, Choice, and Score questions over screenshots, photos, video, and text.

See highlights

Open github.com

Stars: 156
Language: Python
License: APACHE-2.0
Published by: Community
Author: OmniJev
Last verified: Oct 2026

OmniJev/OneJev

Organization repository on GitHub

View on GitHub

🚀🚀 A multimodal System One decision model that gives calibrated answers to typed questions about screens, photos, video and text in one forward pass.

Forks: 17
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

OneJev is a multimodal System One decision model. It returns calibrated probabilities for typed questions about screenshots, photos, videos and text in a single forward pass. Available in 0.8B, 4B, 9B and 27B. Accuracy (%). The OneJev test set is held out from OneJev training. Jev 1.13 uses published text-only scores; Jev-Omni and Qwen3.8-27B thinking were evaluated by us. Choose one backend, then run the Python example below. For NVIDIA GPUs. Supports text, images and video.

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

Tagscommunity jevbest catalog-import

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

PyTorch backend handles text, images, and video on NVIDIA GPUs; the lighter path covers text and images.
Serves a TypeSafe-compatible local API with a Python client for state, media, and batched questions.
Publishes most training questions with media plus configs for one-epoch fine-tuning.
Fine-tuning freezes the vision tower and uses cross-entropy plus Brier loss with sharding for larger sizes.

Watch out

Model choice, hardware, and media size affect latency and accuracy, so confirm behavior on your own screenshots and video. 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.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

intikhab49/open-jev-typed-decision-enginePython

TypeLLM

941

Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

TypeLLM/TypeLLMPython

solar-mini4-jev

Drop-in service that exposes Upstage Solar models through the System One shape, accepting state plus noul, choice, or score questions.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

hunkim/solar-mini4-jevPython

open-medical-jev

Research recipe that fuses two frozen open readers into Jev-style yes/no judgments with a fit-free router, confidence gate, and candidate sets.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

FeiLiuEM/open-medical-jevPython

jev-robot-control

Simulation study that drives one xArm7 apple-to-plate task with Jev, a large reasoning model, and a small model through identical incremental controls.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

openroboto-ai/jev-robot-controlPython

OpenJev-Vision

Open research toolkit that encodes an image once and answers multiple typed choice, noul, and score questions from a shared distribution, plus a local 0.6B text-decision model.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

IamBusy/OpenJev-VisionPython

From the Jev guides

Original write-ups on what this entry is for.

Per-Tick Decisions in Games

Sep 28, 2026

A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.

Jev Demos & Experiments#gaming#agents#control

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai