Jev-Omni

Jev-Omni

Answers a question about text, an image, audio or video by giving a probability for each option you list, instead of writing an explanation. A 12B classifier built on Gemma 4 for developers who need scored decisions.

Made by akhilaaa3 on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Jev-Omni is a decision classifier that works over text, pictures, sound and video. You give it a situation, a question and a list of options. It returns a probability for each option and writes no explanation. It is built on Gemma 4 12B IT and was fine-tuned on 30,000 questions.

Who it suits: Developers who need fast, scored answers to multiple-choice or yes and no questions about mixed media. It fits routing, checking and triage jobs where a probability matters more than prose.

Industry | Who uses it | How they use it | Example
Software | Developers | Add a scored decision step to a pipeline | Ask whether a meeting has started, given the time
Media review | Content teams | Classify clips, images or audio against a question | Pick the best-fitting label for a short video
Operations | Workflow builders | Route items using probabilities rather than text | Choose among a few queues for an incoming item
Research | Evaluation researchers | Compare decision benchmarks | Test it on DecisionBench or JevBench, which the card reports

You can skip it if you want written answers, chat or explanations. It also lacks a CPU path, since the card says a CUDA GPU is required.

Where to use it:

• Python on a machine with a CUDA GPU, with ffmpeg for audio
• Pipelines that call a classifier once per question
• Research setups that compare decision classifiers

Use case | Who | What they get back
Yes or no checks | Developers | A probability for yes and for no
Multiple choice | Workflow builders | A probability for each of up to 20 options
Image, audio or video questions | Media teams | Scored answers about a file you attach

How to set it up: A CUDA GPU is required. The FP32 weights take about 50 GB before runtime overhead, and inference uses BF16 autocast.

• Python and a CUDA GPU
• The packages from the requirements.txt file in the model repository
• ffmpeg, if you plan to send audio

1. Install the packages from the repository's requirements.txt file.
2. Install ffmpeg if you need audio input.
3. Download the repository with huggingface_hub.
4. Load the classifier with the snippet below.

pip install -U huggingface_hub
hf download akhilaaa3/Jev-Omni

How to use it: Download the repository, load the classifier and call predict with a state, a question and your options. The result holds a probability for each option.

from huggingface_hub import snapshot_download

path = snapshot_download("akhilaaa3/Jev-Omni")
import sys; sys.path.insert(0, path)
from jev_omni import load_jev_omni

classifier = load_jev_omni()
result = classifier.predict(
    state="The meeting starts at 10 AM. It is now 9 AM.",
    question="Has the meeting started?",
    options=["Yes", "No"],
)
print(result)

To ask about an image, audio file or video, add a media path and set modality to image, audio or video. Audio is capped at 30 seconds, and video uses 16 frames.

Tool | What it does
load_jev_omni | Loads the classifier and fetches the original Gemma 4 multimodal parts
predict | Takes a state, a question and options, and returns a probability per option
media and modality | Attach an image, audio file or video to the question
snapshot_download | Fetches the repository files, including the loader module

Example: You give it a note saying a meeting begins at 10 and the time now is 9, ask whether the meeting has started, and offer Yes and No. It returns a probability for each, with No scoring high.

For newcomers and experts: Think of it as a quick judge. You describe a situation, ask a question and list the possible answers, and it tells you how likely each one is. Start with the text example and read the printed result.

The model is a fine-tune of Gemma 4 12B IT. On a warm H200, medians over 20 requests were 83 ms for about 2k tokens of text and 26 ms for an image. Thirteen seconds of audio took 31 ms, and a 16-frame video took 504 ms. Calibration error on the medium set is 0.0400, and quality is best with 20 options or fewer.

Supply a state, a question and options
Attach an image, audio or video if needed
The classifier scores every option
Read the probability for each option

Title: Reported scores by benchmark
DecisionBench Medium: 87.57%
JevBench: 86.15%
MMAU: 63.10%
MVBench: 53.10%

Good to know:

• The head accepts up to 256 options, but quality above 20 is not established.
• The card says it is independent and not affiliated with TypeSafe AI or its Jev model.
• Compared models in the card are scored by their developers and may use different protocols.
• License: Apache-2.0, following Gemma 4, with dataset rights kept separate.
• Cost to run: none named, but a CUDA GPU is required.
• Data it reaches: the text, images, audio and video you pass in, plus files downloaded from Hugging Face.
• Not for: written explanations, since it returns probabilities only.

On Hugging Face, 13 November 2026

• Task: text classification
• Library: transformers
• Parameters: 12B
• Weights: about 23.9 GB in BF16; about 6 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 3,744
• Likes: 450
• License: apache-2.0
• Based on: google/gemma-4-12B-it
• Quantized versions: 4
• Fine-tunes: 3
• Tags: transformers, safetensors, gemma4_unified, image-text-to-text, text-classification, multimodal, merged, decision-model

Links

• Model: https://huggingface.co/akhilaaa3/Jev-Omni
• Files: https://huggingface.co/akhilaaa3/Jev-Omni/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download akhilaaa3/Jev-Omni

Then follow the usage in its model card. Nothing in this kit runs by itself.