Jev-Omni
Open multimodal decision model fine-tuned from Gemma 4 12B that returns option probabilities for text, image, audio and video questions.
Overview
Jev-Omni is an Apache-2.0 multimodal decision classifier on Hugging Face, built on Gemma 4 12B IT and fine-tuned on 30,000 questions. You supply a question and options, plus optional media, and receive a probability for each option rather than generated text. It handles text, images, audio up to 30 seconds and video sampled at 16 frames.
The model card reports 87.57% on DecisionBench Medium, 86.15% on matched JevBench groups, 63.10% micro accuracy on MMAU audio and 53.10% on 14 MVBench tasks, and compares with much larger open models on the latter two. It reports warm H200 latencies of 83 ms for about 2k tokens of text, 26 ms for an image, 31 ms for 13 seconds of audio and 504 ms for video. Running it requires a CUDA GPU; FP32 weights use about 50 GB. The list summary notes 1,402 downloads in its first ten days.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Has the meeting started? | Yes · No |
| Choice | Which option answers the question about this image, audio or video? | — |
How it works
- Download the model with huggingface_hub
- Load it with load_jev_omni
- Call predict with state, question, options and optional media
- Use the returned per-option probabilities
Strengths
- Open weights under Apache-2.0
- Covers four modalities in one model
- No output tokens; one forward pass per question
- Published benchmark results
Caveats
- Needs a large CUDA GPU (about 50 GB in FP32)
- One question per call, so state is re-sent for each question
- Video and audio accuracy well below larger models
Getting started
pip install -r https://huggingface.co/akhilaaa3/Jev-Omni/resolve/main/requirements.txt, then load_jev_omni() and classifier.predict(...)At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- Transformers model (CUDA GPU)
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Multimodal System One: a Gemma-4-12B fine-tune that answers typed questions over text, images, audio and video, at 1,402 downloads in its first ten days.
yibie/awesome-jev
Frequently asked questions
What does Jev-Omni decide?
Which option answers a question about text, an image, audio or video, as probabilities.
Is it related to TypeSafe's Jev?
It is an independent open model following the same decision-only idea, fine-tuned from Gemma 4 12B.
What hardware does it need?
A CUDA GPU; FP32 weights use about 50 GB before overhead.
Related projects
Free hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaComfyUI nodes for Jev-style noul, choice and score decisions over images and text using a local imajev adapter on Qwen3.5.
ProjectbetaPythonResearch code for an image-and-text System One model that reads option-letter logits from a multimodal LLM, with training and calibration.
ProjectresearchPythonIndependent open model on Molmo2-ER that answers spatial questions by choosing from fixed options with one shared classifier head.
ProjectresearchPythonOpen project training a vision-enabled Jev-like model (0.8B, 9B) that judges dynamic candidate sets over images with Choice, Noul and Score.
ProjectresearchPythonIndependent Jev-style server on DiffusionGemma and SGLang that classifies doodles, flowers and messages with Choice, Score and Noul.
ServiceresearchPython