OmniJev
OmniJev: omni-modal 4B decision model answering typed questions about images, video, screens and robot scenes in one pass.
Overview
OmniJev, from Beijing Zhongguancun Academy, the Institute of Automation of the Chinese Academy of Sciences and Zevo, is an omni-modal System One decision model. You give it a state of images, video frames, screenshots or camera feeds plus optional text, and typed questions; it returns a probability distribution for every question in one forward pass with zero generated tokens. Choice questions include an abstain probability for none of the above, noul gives a calibrated probability, and score gives a level with probabilities.
It fine-tunes Qwen3.5 vision-language backbones with LoRA, decision heads and an ordinal head. Version 1.1 adds multi-image training and inference, three sizes and an SFT baseline, and covers web and phone operation, robot scenes, video, games, board games, audio spectrograms and text. The README is candid that released demos are offline replays and that reliable closed-loop gameplay and robot control are not established, and that the training-data count is still being audited. Weights are on Hugging Face with an online demo.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which of these options applies to the scene, or none of them? | — |
| Yes / no | Is this statement true of the state? | true · false |
| Score | How far along an ordered scale? | — |
How it works
- Provide images, frames or screenshots plus context.
- Ask typed questions with candidate options.
- Read per-option probabilities and abstain mass.
- Act on high-confidence answers; abstain when none fits.
Strengths
- Native abstention for none of the above.
- Multimodal, multi-image input.
- Open weights and online demo.
Caveats
- Closed-loop control is not established; demos are replays.
- Training-data accounting still under audit.
Getting started
Try it at https://omnijev.net/try.html or download weights from https://huggingface.co/tinnel123/OmniJev.At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- PyTorch
- Has code
- yes
Use cases and integrations
How curated lists describe it
Omni-modal decision model from Beijing Zhongguancun Academy and CASIA: typed questions about images, video, screens and robot scenes, with weights on Hugging Face.
MrJev/awesome-jev
Frequently asked questions
What inputs does OmniJev accept?
Images, video frames, screenshots and camera feeds, with optional text context.
Can it say none of the above?
Yes, choice answers include an abstain probability.
Can it control robots?
The README says reliable closed-loop robot control has not been established.
Related projects
Free hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaIndependent open model on Molmo2-ER that answers spatial questions by choosing from fixed options with one shared classifier head.
ProjectresearchPythonComfyUI 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.
ProjectresearchPythonRepository serving CLM-8B, a contrastive System One model behind a TypeSafe-compatible API, aimed at humanoid robot decisions.
ProjectresearchPythonOpen project training a vision-enabled Jev-like model (0.8B, 9B) that judges dynamic candidate sets over images with Choice, Noul and Score.
ProjectresearchPython