A compact visual reasoner that thinks before it speaks — literally. It runs an internal chain-of-thought before producing answers, which means it tends to catch its own mistakes on tricky image-based problems. At 4B parameters it's surprisingly capable at parsing charts, documents, and scene descriptions, though it can struggle with fine-grained visual details that larger models handle more gracefully.