Typed answers about an image, read from unchanged Qwen3-VL-4B logits in one forward pass.
Want to make some of these yourself?