MODEL PROFILE · UPDATED SEPTEMBER 22, 2026

Jev-Omni

A decision classifier that reads pictures, sound, and video—not just text.

Jev-Omni borrows Jev's typed-decision idea and extends it past text. Give it a state, a question, and a list of options; it returns a probability for every option.

MULTIMODAL DECISION

Jev-Omni

Question

Should this clip be escalated to a person?

Escalate0.91
Keep automatic0.09

BASE MODEL

Gemma 4 12B IT

PARAMETERS

12B

LICENSE

Apache-2.0

AUTHOR

akhilaaa3

WHAT IT CAN READ, AND HOW FAST

Four modalities, one decision interface

Published medians over 20 requests on a warm H200 with an optimised backend. Preprocessing and network time are extra; a smaller card will be slower.

83 ms

Text

Around 2,000 tokens in the published measurement.

26 ms

Image

One image per request.

31 ms

Audio

Capped at 30 seconds. Needs ffmpeg installed.

504 ms

Video

Sampled as 16 frames.

Treat these numbers as a reference point, not a promise: they come from a warm H200 and exclude media preprocessing and network time.

WHAT JEV-OMNI ACTUALLY IS

A classifier, not a chat model

Jev-Omni is independent work by akhilaaa3, not a TypeSafe model. It uses Google's Gemma 4 12B IT with a classification head, fine-tuned across a 30,000-question decision run and published as a merged model.

The head scores the options you supply and returns a probability distribution. It does not generate an explanation or a stream of output tokens, so your application owns the final interpretation and action.

The multimodal input comes from Gemma 4 itself. The loader pulls the required components on first use. The weights are Apache-2.0; the model card notes separately that the training dataset's rights are not covered by that licence.

PUBLISHED RESULTS

Useful numbers, with the context included

The card reports both an equal-weight average across scenarios or groups and a micro average across all questions.

BenchmarkScopeAccuracyMicro accuracy
DecisionBench Medium80 scenarios / 293 questions87.57%86.01%
JevBench (matched subset)195 groups / 231 decisions86.15%87.45%
MMAU1,000 questions63.10%
MVBench14 tasks / 2,786 questions53.10%53.09%
Do not read the JevBench row as a leaderboard place. It is an 86.15% result on a matched subset; JevBench v1.3.0 uses a different measurement and ranking set.

SIZE AND COVERAGE

What the model card compares

These reference scores are reproduced from the model card. The authors note that the larger models report their own official numbers and may use different evaluation protocols.

ModelParametersMMAUMVBenchModalities
Jev-Omni12B63.10%53.10%Text, image, audio, video
Inkling975B total / 41B active77.20%Text, image, audio
Qwen3.5-397B-A17B397B total / 17B active77.60%Text, image, video

Read honestly, this is a size-and-coverage comparison rather than a quality claim. Jev-Omni trails the larger references on their reported benchmarks, at a fraction of the parameters, and is the only one covering all four modalities here.

RUNNING IT

Bring the classifier to your own GPU

A CUDA GPU is required. The FP32 weights are roughly 50 GB before runtime overhead; inference uses BF16 autocast. Audio input also needs ffmpeg.

Before you start

Plan for a CUDA machine with enough memory for the weights and runtime. Validate accuracy, calibration, and media preprocessing on your own data before putting the output in a critical path.

pip install -r https://huggingface.co/akhilaaa3/Jev-Omni/resolve/main/requirements.txt
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(state, question, options)

For image, audio, or video, add media="/path/to/file" and modality="image", "audio", or "video".

LIMITS WORTH KNOWING FIRST

The trade-offs are part of the model

Jev-Omni is useful when the decision is bounded and multimodal. It is less useful when you need open-ended generation, long media, or a lightweight hosted endpoint.

  • Twenty options, practically. The head accepts 256, but quality above 20 is not established.
  • Short media only. Audio is capped at 30 seconds and video is sampled to 16 frames.
  • It is a classifier. There are no generated output tokens and no built-in explanation for the chosen option.
  • Serious hardware. Expect about 50 GB of weights and a CUDA GPU; published speeds were measured on an H200.
  • Dataset rights are separate from the Apache-2.0 licence on the weights. Check the model card before redistribution.

JEV-OMNI OR JEV?

Choose based on the state you need to understand

The boundary is simple: Jev-Omni adds media input and self-hosting; Jev stays focused on fast, hosted text decisions.

REACH FOR JEV-OMNI IF

The decision is multimodal

  • The decision is about an image, a short voice clip, or a few seconds of video.
  • You have a CUDA GPU with room for a 12B multimodal model and want open weights you control.
  • You are willing to validate accuracy and calibration on your own data before trusting the numbers.

REACH FOR JEV IF

The decision is text

  • Your state is text, which covers most triage, moderation, routing, and scoring work.
  • You want calibrated probabilities from a hosted API with nothing to install and no GPU to maintain.
  • You want to start in a browser in the next minute rather than after a 50 GB download.

The text half can be tested right now

Run a real case in the Jev playground and inspect the full probability distribution before you write an integration.

Open the playground

QUESTIONS

Jev-Omni FAQ

What is Jev-Omni?+

An open-weights multimodal decision classifier by akhilaaa3, fine-tuned from Google's Gemma 4 12B IT on 30,000 decision questions. You give it a state, a question, and options; it returns probabilities.

Is Jev-Omni made by TypeSafe, or related to Jev?+

No. It is independent work that borrows the Jev name and typed-decision idea. TypeSafe's Jev is a closed hosted model; Jev-Omni is Apache-2.0 weights on Gemma 4 that you run yourself.

Does Jev-Omni appear in the JevBench ranking?+

Not as a directly comparable ranking entry. Its card reports 86.15% on a matched subset of JevBench tasks, while JevBench v1.3.0 is a different measurement across a broader set of systems.

What hardware does Jev-Omni need?+

A CUDA GPU. The FP32 weights are about 50 GB before runtime overhead, and inference runs under BF16 autocast. The published latencies were measured on a warm H200, so treat them as a ceiling rather than a guarantee.

How many options can Jev-Omni handle?+

The classification head accepts up to 256, but the card states it is best supported at 20 or fewer and that quality above 20 is not established. Test your own option counts before relying on them.

Can Jev read images the way Jev-Omni does?+

No. Jev is text-only: a string, a JSON object, or an array. If your decision is about a photograph, a voice note, or a clip, use a multimodal model such as Jev-Omni.

SOURCES

Read the primary material

Sources checked September 22, 2026.

Jev AI GitHub

Jev-Omni, Gemma, and every other product named on this page belong to their respective owners. Figures are quoted from the linked sources and are presented for orientation, not as a guarantee of performance.