Jev-Omni — open multimodal typed decisions (text / image / audio / video)
On 21 September 2026, independent author Akhila Datta (akhilaaa3 / @Akhila_988) published Jev-Omni on Hugging Face: an open ~12B System One–style option classifier on google/gemma-4-12B-it that takes state + question + options (optional media= for image/audio/video) and returns {prediction, prediction_index, confidence, probabilities} — no autoregressive text output.
Inference uses a custom loader (jev_omni.py) plus head.pt — not a pure transformers AutoModel drop-in. License Apache-2.0. The card disclaims any TypeSafe / Jev affiliation and states the model was not trained on Jev output. Dataset called out on the card: akhilaaa3/decision-bench.
Catalog framing: first open catalog peer with all four modalities shipped (text/image/audio/video). openjev already has image/video/text; Prosodia remains an audio-specialist off-quadrant. Do not treat the tweet’s “first multimodal system one” line as a category fact.
UnverifiedClaim / hygiene: DecisionBench Medium (87.57% / micro 86.01%), JevBench matched (86.15% / 87.45%), MMAU 63.10%, MVBench 53.10%, and warm-H200 latency rows (83 / 26 / 31 / 504 ms) all come from the HF card. Card “30,000-question” FT vs decision_config recipe size: 24000 vs tweet “30k examples / <100 ms on 1 H100” conflict across sources. The launch X video is animated motion-graphics packaging (bottom disclaimer: real model outputs; choices can be any modality) — on-screen latencies differ from the H200 table, and demo bars can show probs that do not sum to 1. None of these are ModelSystem.One measurements.
Catalog entry: /models/jev-omni. Peer context: openjev, Tiny-Jev, hosted Jev.