Teaching a frozen LLM to see

We gave our production LLM (a fine-tuned GLM-5.2 checkpoint) the ability to read images by training a 49.5M-parameter projector and two marker embeddings, nothing else.