The Allen Institute for AI (AI2) has released MolmoAct 7B, an open-source "action reasoning model" designed to let robots plan movements in three-dimensional space before executing them. The model interprets natural-language commands, lifts a visual scene into 3D, and charts a motion trajectory that developers can preview and adjust before a robot acts. MolmoAct 7B contains seven billion parameters and was trained on 18 million samples using 256 Nvidia H100 graphics processors, with fine-tuning completed on 64 H100s. The dataset includes roughly 12,000 real-world robot episodes from environments such as kitchens and bedrooms. On the SimPLER benchmark, the system achieved a 72.1% task-success rate, outperforming rival offerings from Nvidia, Google, Microsoft and startup Physical Intelligence. Chief Executive Ali Farhadi said the release aims to provide a transparent foundation for embodied AI, while computer-vision lead Ranjay Krishna highlighted the model’s ability to map entire scenes into 3D before taking action. AI2 has published the code, weights and evaluations, positioning MolmoAct as a freely available alternative to proprietary robot-control models.