DownloadJev-Omni is an open-weight multimodal decision classification model designed to evaluate text, images, audio, and video through structured questions and predefined answer options. Built on Google's Gemma 4 12B IT architecture and fine-tuned on 30,000 questions, it returns calibrated probabilities for each possible answer rather than generating natural-language explanations. The model supports three decision formats: Noul for binary yes/no questions, Choice for multiclass selection, and Score for probability-based evaluation. It achieved 87.57% accuracy on DecisionBench Medium, 86.15% on JevBench, 63.10% on MMAU, and 53.10% on MVBench. Jev-Omni is designed for classification, automated decision-making, multimodal assessment, and applications requiring measurable confidence. It supports up to 20 answer options with established quality, processes audio clips up to 30 seconds and video through 16 sampled frames, and runs locally on CUDA GPUs. It is released under the Apache 2.0 license.
Features
- 12B-parameter model based on Gemma 4 12B IT
- Multimodal classification across text, images, audio, and video
- Noul, Choice, and Score structured decision formats
- Calibrated probability outputs without generated explanations
- Fine-tuned on 30,000 decision questions
- 87.57% accuracy on DecisionBench Medium
- Supports up to 20 validated answer options per question
- Local CUDA inference with Apache 2.0 licensing
Project Samples
