Click a stage. Every picture below is the real output of that stage for this page, captured from the model.
Each layer's state for this token, put through the final norm and the output layer early: the word the model would write if it stopped at that layer. Watch the answer form in the last layers.
All of modeling_qwen3_vl.py for one forward pass over this page and the prompt: each card is a module (a layer with weights, or a wrapper around layers), with its real output drawn over the page, its output shape and a picture of its weight matrix. Dashed cards are steps written as plain code, not modules. Hover for the details, click to inspect.