Molmo: the open-weight VLM that beats GPT-4V at pointing

Most VLMs can describe what's in an image, but can't indicate exactly where something is located. Molmo solves this problem: if you ask "point to the glass on the table" it responds by pointing with precise coordinates on the image. Allen AI made public not just the model but also the PixMo dataset used to train it, created with detailed voice descriptions collected from humans. This "full open pipeline" approach is rare and invaluable for research.