The Glance method is presented for fast classification in the Qwen3-VL-4B model

Developers introduced Glance — a specialized data reading method for the multimodal model Qwen3-VL-4B. The tool allows you to get answers in the "Yes/No" format or choose options from a list in one pass through the neural network. In contrast to the standard approach, Glance does not generate text tokens, but directly reads the probabilities of answers. This significantly reduces the processing time of requests: the system shows the result in just 1 hour when performing 14 test iterations. This approach optimizes work with visual data, eliminating the stage of decoding long sequences. Despite the high speed, the current version of the model has technical limitations. During tests, it was found that Qwen3-VL-4B has difficulties in determining the relative sizes of objects in images. Also, the model incorrectly handles tasks for counting objects if their number exceeds five units. The tool is available for testing on the Replicate platform. The method is focused on classification and image analysis tasks, where an instant system response is required without the need to form a detailed text description. The use of probabilistic inference allows integrating the model into pipelines that require a strict structure of answers.