A faster way to ask vision models
Why make a vision model write JSON when all you need is a choice? Today’s build pattern reads the model’s decision scores instead.
In today’s edition
- Glance VLM speeds local visual decisions.
Glance VLM skips generated text for visual choices
For a yes-or-no image check, a model does not always need to write an answer. Glance VLM reads a local vision model’s logits, or decision probabilities, to select from fixed choices without generating JSON.
- The package is presented as Apache-2.0 software that runs locally, with support tested on Qwen3-VL 2B, 4B, and 8B models and SmolVLM2.
- The creator reports a full-photo yes-or-no decision in 1.1 seconds on a laptop, with roughly 1.5x faster results on full images and 2x to 6x gains for many questions on smaller images.
- Reported tests found the Qwen 4B model scored 0.939 when reading logits versus 0.945 when generating text on yes-or-no tasks, and 0.933 versus 0.930 on pick-one tasks.
- This is not a cure-all: the reported results show weak geometry performance, including 0.52 on relative size and 0.65 on diagonal direction.
What it means For repeated yes-or-no or fixed-choice image checks, test logit reading before text generation to keep images local and cut latency, formatting failures, and GPU spend.
Source: yoheinakajima