- Category
- AI Tools
- Rank
- No. 3772Tools index
Previous survey · No. 3574 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Models: Train & Run
- GitHub
- 34 stars
- Latest release
- v0.2.1
- Date
About
Vev is a set of open-weight decision models built on Qwen3.5 that can read images as well as text and JSON. You ask yes/no, multiple-choice or graded questions about a screenshot, and it returns a probability for every allowed answer in about 40 ms. Nothing is generated, so answers are always one of the options you define and you can apply a threshold to them.
At a glance
Vev is a research model for answering fixed questions about an image or text and returning scores for the allowed answers. It is worth evaluating for non-commercial screenshot triage when a workflow needs a label and a way to route uncertain examples for review.
Good fit
Researchers with a GPU and labeled examples for tasks such as detecting an error state in a screenshot.
Consider something else when
Pass for commercial deployment under the published weights license or when the task needs a free-form explanation.
Start here
Score a labeled set of synthetic screenshots with a fixed option order. Compare false negatives at several thresholds before letting a score trigger any action.
Cost and requirements
The code is Apache 2.0 and the weights are CC BY-NC 4.0. Local inference requires model-serving compute, with memory needs depending on the checkpoint and input length.
Limits to know
The maintainers report sensitivity to answer order and uneven calibration. Treat scores as rankings until you have measured them on your own data.
Sources checked 2026-11-05. Source-based assessment.
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.