Jev-Omni scores fixed choices across four modalities

Jev-Omni is a 12-billion-parameter open-weight classifier that accepts text, images, audio, or video and assigns a probability to each supplied answer. A request contains a state, meaning the context being evaluated, along with a question and candidate options. The fixed output contract gives applications structured decisions without parsing generated prose.

The model was fine-tuned from Gemma 4 12B IT on 30,000 questions. Its author claims it is the first downloadable checkpoint to support all four modalities through one decision interface, a combination that many open-weight alternatives lack. The repository lists an Apache-2.0 license; adopters should also review the terms attached to the upstream Gemma components downloaded by the loader.

Fixed choices, usable probabilities

Generative multimodal models emit token sequences that structured pipelines must validate and parse. Jev-Omni directly scores the options supplied with each question, reducing the need for regular expressions, constrained decoding, or a separate judge model.

The typed-decision interface supports yes/no, multiple-choice, and score questions. On the medium split of DecisionBench, the model card reports an expected calibration error of 0.0400 across 10 confidence bins. Expected calibration error measures the weighted gap between predicted confidence and observed accuracy, so 0.0400 represents an average gap of about four percentage points on that test split. Calibration can change with new data and deployment conditions.