Ai2's Open Robot Model Beats the Closed Competition

The Allen Institute for AI has released MolmoAct 2, an open-source foundation model for robots, and the claim attached to it is the kind that usually comes wrapped in a closed API and a sales call: it outperforms capable proprietary robotics models on industry benchmarks, without per-task fine-tuning. Ai2 published the model, its weights, and its training recipe in full.

That openness is the point. Most of the impressive robotics work of the past two years has happened inside a handful of well-funded labs, behind walls. Ai2 has built much of its reputation on doing the opposite, and MolmoAct 2 is the strongest argument yet that an open model can sit at the frontier rather than trailing it.

A robot that thinks before it acts

What sets the model apart is how it approaches a task. Ai2 calls the design an Action Reasoning Model: rather than mapping a camera feed straight to motor commands, MolmoAct 2 reasons about the three-dimensional scene in front of it before deciding what to do. The difference shows up in generality. Straight from the weights, the model can move laboratory objects, handle scientific tools and fold a towel, tasks that normally demand careful per-robot tuning.

It is also fast. Ai2 says MolmoAct 2 runs about 37 times quicker than the original MolmoAct, the gap between a research curiosity and something you could imagine running on real hardware in real time.

The data nobody else is sharing

Alongside the model, Ai2 released what it says is the largest open bimanual tabletop manipulation dataset ever published: more than 720 hours of two-handed robot demonstrations. In a field where the training data is often the real moat, handing that over matters as much as the model itself. The institute also open-sourced the action tokenizer that turns the model's decisions into robot motion.

Put together, it is an unusually complete release. A team that wants to build a physical AI system now has the model, the data and the tooling to start, rather than a paper describing results they cannot reproduce.

Why this lands now

Physical AI, the effort to give robots the kind of general competence that language models brought to text, is where a lot of the field's attention is heading. The big question is whether that future gets built in the open or licensed out piece by piece. MolmoAct 2 is a deliberate push toward the former, and it follows a string of capable open releases, including Zyphra's ZAYA1, that keep narrowing the distance between open and closed AI.