MolmoMotion

"Forecasting Point Trajectories in 3D with Language Instruction": a 4.85B Molmo2-architecture VLM that predicts 3D point trajectories from RGB video plus natural-language instructions, transferring directly to robot pick-and-place. Two released variants (H3-F30, H1-F32) plus the 1.16M-video molmo-motion-1m training set and the PointMotionBench evaluation. Apache-2.0. Extends the Molmo line from perception into motion forecasting for robotics.