A PDF is read from its first page. The GPU holds one model at a time and swaps the other in within seconds; the first run after a pause waits a few minutes while the GPU starts and loads the weights. Everything the prepared prompts show is captured for your document too, and every tab can then show it.
No password? Prepared prompts has twelve documents and questions already run on both models, with every view.
Loading the runs…
Every attention layer of both models, one at a time: move through the layers and through the answer. Qwen has 16 full-attention layers (its other 48 layers are Gated DeltaNet, with no attention map: see DeltaNet memory). Gemma has 30: 25 that only see the last 1,024 tokens and 5 global ones.
At every layer, the model's final norm and output head can be applied early: what would the model write if it stopped here? Play the layers and watch the answer form. Green: the token this layer would write is the one the model finally wrote.