Text
Around 2,000 tokens in the published measurement.
83 msModel profile · updated September 22, 2026
Jev-Omni
A decision classifier that reads pictures, sound and video — not just text.
Jev-Omni is an independent, open-weights model that borrows Jev's idea and extends it past text. You hand it a state, a question and a list of options; it hands back a probability for each option and writes nothing. What makes it worth knowing about is the input side: it also accepts images, audio and video, which Jev does not.
- Base modelGemma 4 12B IT
- Parameters12B
- LicenceApache-2.0
- Authorakhilaaa3
What it can read, and how fast
Warm H200, medians over 20 requests on an optimised backend. Preprocessing and network time are extra, and a smaller card will be slower.
Image
One image per request.
26 msAudio
Capped at 30 seconds. Needs ffmpeg installed.
31 msVideo
Sampled as 16 frames.
504 msWhat Jev-Omni actually is
Independent work by akhilaaa3, not by TypeSafe.
It is Google's Gemma 4 12B IT with a classification head bolted on, 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 distribution over them. No tokens are generated, so there is no prose to parse and no schema for the model to break.
The multimodal parts come from Gemma 4 itself: the loader pulls the original multimodal components down automatically the first time you use them. Audio is capped at 30 seconds and video is sampled down to 16 frames, so this is a model for short clips and single images rather than long recordings.
It is Apache-2.0, following Gemma 4's terms. The card notes separately that rights to the training dataset are not covered by that licence.
Published results
Both columns the card reports: an equal-weight average across scenarios or groups, and a micro average across all questions.
| Benchmark | Scope | Accuracy | Micro accuracy |
|---|---|---|---|
| DecisionBench Medium | 80 scenarios / 293 questions | 87.57% | 86.01% |
| JevBench (matched subset) | 195 groups / 231 decisions | 86.15% | 87.45% |
| MMAU | 1,000 questions | — | 63.10% |
| MVBench | 14 evaluated tasks / 2,786 questions | 53.10% | 53.09% |
Do not read that JevBench row as a leaderboard place. The card reports 86.15% on a matched subset — 195 groups and 231 decisions. Benchmark Heaven's JevBench v1.3.0, the ranking used on the comparison pages of this site, is a different measurement: 534 decisions across 52 systems, scored on a composite of intelligence, calibration, speed and cost. Jev-Omni does not appear in it. An accuracy on a subset is not a rank, and the two numbers are not comparable.
Calibration: the card reports a medium expected calibration error of 0.0400 over 10 bins, measured by its author.
How the card positions it against larger open models
Reproduced from the model card, which notes that the reference scores are officially reported by their own developers and may use different evaluation protocols.
| Model | Parameters | MMAU | MVBench | Modalities |
|---|---|---|---|---|
| Jev-Omni | 12B | 63.10% | 53.10% | Text, image, audio, video |
| Inkling | 975B total / 41B active | 77.20% | — | Text, image, audio |
| Qwen3.5-397B-A17B | 397B total / 17B active | — | 77.60% | Text, image, video |
Read honestly, this table is a size argument rather than a quality one: Jev-Omni trails both references on their own benchmarks, at a fraction of the parameters, and is the only one of the three covering all four modalities.
Running it
A CUDA GPU is required. The FP32 weights are roughly 50 GB before runtime overhead; inference uses BF16 autocast. Audio input also needs ffmpeg.
pip install -r https://huggingface.co/akhilaaa3/Jev-Omni/resolve/main/requirements.txtfrom 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)For other modalities, add media="/path/to/file" and modality="image", "audio" or "video".
Limits worth knowing first
- Twenty options, practically. The head accepts 256, but the card says quality above 20 is not established.
- Short media only. 30 seconds of audio, 16 frames of video.
- It is a classifier. No generated output tokens, so no explanation of why it chose what it chose.
- Serious hardware. About 50 GB of weights and a CUDA GPU, with published speeds measured on an H200.
- Dataset rights are separate from the Apache-2.0 licence on the weights.
Jev-Omni or Jev?
Reach for Jev-Omni if
- 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
- Your state is text, which is most triage, moderation, routing and scoring work.
- You want calibrated probabilities from a hosted API with nothing to install and no GPU to keep running.
- You want to start in a browser in the next minute rather than after a 50 GB download.
The text half of this you can test right now
If your decisions are text, the Jev AI playground runs Jev 1.13.0 on your own cases with five free credits, no setup and the full probability distribution on every answer. If they are not text, Jev-Omni is a genuine option and Jev is not — convert to text or use a model that takes pixels.
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 a list of options; it returns a probability for each option and generates no text.
Is Jev-Omni made by TypeSafe, or related to Jev?
No. It is an independent model that borrows the Jev name and the 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?
No. Its card reports 86.15% on a matched subset of JevBench tasks — 195 groups and 231 decisions. Benchmark Heaven's JevBench v1.3.0 is a different measurement: 534 decisions across 52 systems, scored on a composite of intelligence, calibration, speed and cost. An accuracy on a subset is not a rank, and the two numbers cannot be compared.
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 what you would see on a smaller card.
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, Jev cannot make it and a multimodal model like Jev-Omni can.
Sources
Read on September 22, 2026.
Jev-Omni, Gemma and every other product named on this page belongs to its respective owner and is not affiliated with Jev AI. Figures are quoted from the sources above as they read on September 22, 2026 and can change.