MolmoPoint is also easier to train—with just 8K examples it outperforms coordinate-based models by ~20 F1 points & reaches peak performance faster during full pretraining.
These grounding gains don't come at a cost. Question-answering, captioning, & other tasks all stay on par.