Ai2 Releases MolmoAct 2 Open Robotics Foundation Model

The Allen Institute for AI released MolmoAct 2, a fully open robotics foundation model that runs 37x faster than its predecessor and outperforms Physical Intelligence's π0.5 on bimanual manipulation benchmarks.

Ai2 Releases MolmoAct 2 Open Robotics Foundation Model

Ai2 MolmoAct 2 open robotics model

The Allen Institute for AI (Ai2) has released MolmoAct 2, the second generation of its open robotics foundation model, along with the largest open-source dataset for two-armed tabletop robot manipulation. The release, announced May 5, 2026, ships with full open weights, training code, evaluation rollouts and a 700-hour bimanual manipulation dataset called MolmoAct 2-Bimanual YAM.

What's new in MolmoAct 2

Ai2 says MolmoAct 2 runs up to 37x faster than its predecessor and outperforms Physical Intelligence's π0.5 in simulations, zero-shot real-world tasks and third-party evaluations. The model is designed for everyday manipulation: it can move objects, pour, fold and handle laboratory tools without task-specific retraining. The architecture is a vision-language-action model that emits low-level motor actions directly from camera and text input.

Why open weights matter

The release of full training data and code is unusual in the robotics-foundation-model space, where rivals including Physical Intelligence and Skild AI keep their datasets proprietary. Ai2 has been pushing the open-source agenda hard, having previously released the OLMo language family and the original Molmo multimodal models. Researchers can now reproduce the MolmoAct 2 training run end to end, and hardware vendors can fine-tune it for new robot embodiments without negotiating commercial licenses.

Robotics ecosystem reaction

The release lands alongside an accelerating push by NVIDIA's Isaac and GR00T stack to standardize humanoid and manipulation reference platforms. Ai2's bet is that an open, fast, real-world capable model will catalyze a broader contributor base, much as Llama did for language models. Early integrations are expected from academic and startup labs that want a competitive base model without the cost of bootstrapping their own dataset.

Available now

MolmoAct 2 model weights, the YAM dataset, training scripts and a fully reproducible recipe are available on the Ai2 site and Hugging Face. The Seattle non-profit said additional task suites and a video-action variant called Molmo 2 are scheduled for later this year.

Reporting based on coverage from SiliconANGLE, The AI Insider, TechTimes and Ai2's official MolmoAct 2 blog post.

Category: Machine Learning

Tags: Open Source AI Machine Learning AI Models robotics startup AI computer vision

Related Articles