Glance
Calibration harness that reads yes/no, pick-one and rating answers from one forward pass of a frozen open vision-language model.
Overview
Glance (glance-vlm) lets you ask an open vision-language model typed questions about an image and get probabilities back on your own machine. Yes/no, pick-one and rating answers are read from the logits of one forward pass of a frozen open model, Qwen3-VL-4B by default; nothing is generated or trained, and no image leaves the machine. The README stresses that Glance is a calibration and measurement harness around that readout, not a model.
It offers a CLI and Python API mirroring Jev's primitives (noul, choice with options, score with levels), a doctor command that picks a model for your hardware (Apple silicon or CUDA, with slow CPU fallback), and stores logs and fitted calibrations locally. About 9 GB of open weights download on first use. A project page publishes every result, latency experiments and a local camera lab.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is there a dog in the image? | yes · no |
| Choice | What is the main subject? | dog · cat · car |
| Score | How blurry is the photo? | Sharp · Slightly soft · Blurry · Very blurry |
How it works
- pip install glance-vlm and run glance doctor
- Ask a typed question about an image
- Read probabilities from one forward pass
- Fit and apply calibration from your labels
Strengths
- Fully local; images never leave the machine
- No training; frozen open model
- Calibration harness included
Caveats
- About 9 GB download
- CPU inference is slow
Getting started
pip install glance-vlm, glance doctor, then glance ask photo.jpg "Is there a dog?".At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- CLI + library (local VLM)
- Has code
- yes
Use cases and integrations
How curated lists describe it
Asks a frozen open vision-language model typed questions about an image and reads the answer out of one forward pass, with a calibration harness around the readout.
MrJev/awesome-jev
Frequently asked questions
Is Glance a model?
No, it is a calibration and measurement harness around an open VLM's readout.
Do images leave my machine?
No.
Which model is used by default?
Qwen3-VL-4B.
Related projects
Research fork of Laya that reads typed choice and yes/no decisions from an image plus optional text in one forward pass.
ProjectresearchPythonDecision-forensics library that explains a Jev answer by ablating parts of the input and measuring how far the probability moves.
ProjectbetaPythonAdapter that turns existing LLMs into Jev-style decision functions with option scores, optional calibration and review thresholds.
ProjectprototypePythonDecision sidecar and Copilot CLI pre-turn hook that mimics Jev choice, noul and score using Azure Foundry Cohere Rerank.
HookprototypePythonResearch code for an image-and-text System One model that reads option-letter logits from a multimodal LLM, with training and calibration.
ProjectresearchPythonExact-target audit finding Jev's Choice probabilities are a step function on stated base rates while per-option Noul tracks them closely.
ProjectresearchPython