Model guide · updated September 25, 2026

OmniJev

A visual decision model for screenshots, images and sampled video. OmniJev scores the questions you supply, returning structured probabilities instead of writing a response.

Looking for the Gemma-based model? Read the Jev-Omni guide. OmniJev uses Qwen3.5; Jev-Omni uses Gemma 4 12B. They are separate projects. Compare OmniJev vs Jev-Omni.

Try native Jev in the playground How to use Jev

OmniJev is not available through this site’s API. The playground button opens native Jev for text decisions.

Capability snapshotACCURACY

Six selected tasks · 0–100% per axis.
Author-reported results, not an overall score.

3 sizes0.8B · 2B · 4B

Visual inputImages & screenshots

16 framesSampled per video

0 tokensStructured decision output

Measured, not inferred

Published accuracy

Held-out task results reported by the OmniJev authors. Each bar spans 0–100%; higher accuracy is better. Qwen3.5-4B is the zero-shot backbone baseline.

OmniJev accuracy by model size · September 25, 2026
Benchmark / samples	OmniJev-0.8B	OmniJev-2B	OmniJev-4B	Qwen3.5-4B
LIBERO-10 robot decisionsn = 1,504	77.1%	72.4%	80.7%	29.9%
Mind2Web test (task / website / domain)n = 1,500	59.6%	63.1%	73.3%	30.3%
Charades-STA video eventsn = 1,500	81.1%	82.6%	85.9%	54.3%
OK-VQA answer pooln = 1,500	65.6%	76.5%	80.9%	86.0%
LongVideoBench valn = 500	48.6%	50.2%	58.2%	58.5%
ESC-50 sounds (spectrogram)n = 795	47.3%	51.4%	52.5%	34.7%
Super Mario Brosn = 735	—	—	33.9%	46.4%

Performance varies by task: the 4B release trails its backbone on Super Mario Bros and slightly on LongVideoBench. A dash means no result was reported. These are not comparisons with Jev-Omni.

What is OmniJev?

The project credits Beijing Zhongguancun Academy, the Institute of Automation of the Chinese Academy of Sciences, and Zevo. Its released models use Qwen3.5 backbones and a decision interface that answers several typed questions about a shared visual state.

The repository demonstrates browser and phone actions, robot observations, game decisions, gesture recognition and video-event judgments. These are examples of scoring a supplied state and candidate decisions; an application still needs to choose thresholds and execute any resulting action.

Available OmniJev models

Download both the decision checkpoint and its matching backbone. The project states Apache-2.0 for its code and weights; the backbone retains its own license.

OmniJev-4B

Qwen3.5-4B

tinnel123/OmniJev

OmniJev-2B

Qwen3.5-2B

tinnel123/OmniJev-2B

OmniJev-0.8B

Qwen3.5-0.8B

tinnel123/OmniJev-0.8B

Three question types

choice scores named candidates and includes an abstention probability. Candidates can also identify image regions.
noul estimates whether a statement holds for the supplied state.
score evaluates an ordered set of levels, such as stages of task completion.

Questions are keyed by ID, making it possible to associate each answer with the original request. Sharing one visual state across several questions avoids repeating the entire image-processing step for each question.

Getting started

Follow the repository's environment setup, then download a matching checkpoint and backbone. This example uses the 4B release.

hf download tinnel123/OmniJev --local-dir ckpt
hf download Qwen/Qwen3.5-4B --local-dir base

from mso.infer import MSO1

model = MSO1("ckpt", "base")
answers = model.system_one(
    {"images": ["screen.png"]},
    {"error_visible": {
        "type": "noul",
        "instructions": "An error dialog is visible on the screen."
    }}
)
print(answers["error_visible"])

Video preparation uses the repository's video_state helper and requires ffmpeg and ffprobe. Sampling a clip into frames can miss brief events between those frames.

Before using it in an application

A probability is a model estimate, not a guarantee that an action is correct. Evaluate your own screenshots, candidate choices and abstention thresholds before connecting decisions to a browser or robot.

The repository provides task-specific accuracy, calibration and latency measurements. Those are author-reported results; hardware, image preparation and question count affect runtime. Smaller model sizes alone do not establish the memory budget for your deployment.

This guide does not add OmniJev to the Jev AI API or playground. For the hosted text-decision workflow available here, see What is Jev? or the Jev API guide.

OmniJev FAQ

What does OmniJev return?

It returns structured decisions: choice probabilities with an abstain option, a probability for a noul statement, or probabilities over ordered score levels. It does not generate a written explanation.

Can OmniJev process audio?

The repository reports sound classification using spectrogram images. Its documented image and video workflow should not be treated as a general audio-file interface.

Can I use OmniJev in the Jev AI playground?

This page is a model guide, not an announcement of a hosted OmniJev integration. Use the project demo or follow its repository instructions to run the released weights.

Sources & model weights · reviewed September 25, 2026

OmniJev belongs to its creators. This page is an independent guide; project capabilities and releases may change.