MolmoAct 2

MolmoAct 2 is an open AI robotics foundation model from the Allen Institute for AI (Ai2), built as an Action Reasoning Model (ARM) — a class of models that reason about their environment in 3D before directing robot actions. It brings faster, stronger 3D action reasoning to real-world robot tasks and handles bimanual (two-handed) manipulation without requiring per-task fine-tuning. According to Ai2, MolmoAct 2 runs up to 37x faster than the original MolmoAct and is released together with a major new bimanual manipulation dataset for researchers to study, reproduce, and build on. It is fully open: Ai2 has published the model, a technical report, all training data, fine-tuning scripts, evaluation rollouts, and the training recipe for the open-source tokenizer, so anyone can adapt it to new hardware or extend it to different tasks. MolmoAct 2 is also integrated into Hugging Face's LeRobot platform, letting teams drop it into existing setups without retooling, and Ai2 reports its artifacts have been downloaded more than 400,000 times. Aimed at robotics researchers and ML engineers, MolmoAct 2 targets real-world tasks — like loading a dishwasher or prepping lab samples — where reliable, reproducible robot autonomy has been hard to achieve.

Robotics researchers and ML engineers use MolmoAct 2 to train and study robots for real-world 3D manipulation tasks.