How a Multimodal LLM Sees

Step through the full pipeline: text is tokenized, the image is cut into patches, both become vectors in a single shared sequence, the transformer runs cross-modal attention, mixture-of-experts routing and feed-forward activations, and the model produces a next-token distribution. Every value is derived live from the pixels and the text you type - illustrative numbers, real architecture.