Glance readout for Qwen3-VL-4B
This page runs the unchanged Qwen3-VL-4B-Instruct weights with the Glance readout. Glance is not a model or a fine-tune. It is a way to ask an open vision-language model a typed question about an image and read probabilities from the model’s logits in one forward pass. Nothing is generated.
This Cog package installs glance-vlm==0.3.1, constructs the package’s default Glance() configuration, and uses its noul and choice readouts.
What it returns
- Yes / No: a probability for Yes and No, with the higher-probability answer selected.
- Multiple choice: one probability for each option you provide, with one of those options selected.
A chat model normally writes an answer one token at a time and the result then has to be parsed. Glance skips the writing. It runs the model once and reads the scores assigned to the allowed answers.
Limits
The answer is always one of the allowed options, but it can still be wrong. The probabilities are not calibrated: a value near the middle is a useful flag, while a high value is not a guarantee. This 4B model is weak on relative size, the direction of diagonal lines, and counting past five.
Images are sent to Replicate to run the model, and Replicate’s data retention policy applies. To keep images on your own machine, run the local package.