Vision-Jev
Open project training a vision-enabled Jev-like model (0.8B, 9B) that judges dynamic candidate sets over images with Choice, Noul and Score.
Overview
Vision-Jev is an open foundation project for judging and acting over a dynamic candidate set. Given an image, visible state, a question and changing candidates, the model produces a Choice (select or rank the best candidate), a Noul (whether evidence is sufficient or a proposition is supported) and a Score (an ordered quality or confidence level). The same visual representation is meant to support a separate policy and value branch for closed-loop interaction.
Checkpoints at 0.8B and 9B are compared with the original Qwen3.5 on frozen test samples covering visual question answering, text reading, chart and diagram reasoning, evidence sufficiency and visual quality. The project commits to publishing the full training path: data pipeline, quotas, manifests, SFT and policy-training code, evaluation protocol and evidence. It is Apache-2.0 with weights on Hugging Face.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which candidate best answers the question about this image? | — |
| Yes / no | Is the visual evidence sufficient to support this proposition? | yes · no |
| Score | What is the ordered visual quality level? | — |
How it works
- Provide an image, state, question and candidates.
- The model scores candidates, evidence sufficiency and quality.
- Apply frozen confidence thresholds.
- Optionally use the policy/value branch for interaction.
Strengths
- Open data-to-weights development.
- Handles changing candidate sets.
- Apache-2.0.
Caveats
- Training path partly promised rather than complete.
- Benchmarks are the project's own.
At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- PyTorch
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
An open vision-enabled JEV-like model.
hellogumbo/awesome-jev
Frequently asked questions
What sizes exist?
0.8B and 9B checkpoints.
What does it decide?
Best candidate, evidence sufficiency and quality level over images.
Is the training open?
The project commits to publishing the full data and training path.
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.
ProjectresearchPythonIndependent Jev-style server on DiffusionGemma and SGLang that classifies doodles, flowers and messages with Choice, Score and Noul.
ServiceresearchPythonCloudflare's Apache-2.0 27B multimodal decision model that scores every option of every question jointly, with a Jev-compatible API.
ServiceproductionPython