Updated on
In August 2025 Ai2 released MolmoAct, an open model that reasons in 3D; MolmoAct 2 followed in 2026. What held up and what to doubt.
In August 2025 the Allen Institute for AI (Ai2) released MolmoAct, a 7-billion-parameter model it calls an Action Reasoning Model (ARM): instead of going straight from words to motor commands, it first reasons about the scene in space. Models, code and data were published openly, and the weights sit on Hugging Face. Over a year later, with a successor out, it is worth separating what was new from what was marketing.
The pitch is a critique of the usual recipe. Vision-language-action models lean on language, which Ai2 argues is a poor carrier for motion and physics. MolmoAct instead grounds the scene with depth-aware perception tokens, sketches a visual reasoning trace, then outputs actions. In plain terms: it plans in image space before it moves, which also makes the plan something a human can inspect.
Now the numbers, with the caveat that they are Ai2's own. The company reports 72.1% out-of-distribution success on SimplerEnv and an 86.6% average on LIBERO after parameter-efficient fine-tuning, and compares against models from Physical Intelligence, Google, Microsoft and NVIDIA. The Robot Report quoted slightly different figures (71.9% on SimPLER, about 12,000 robot episodes, 18 million samples) than Ai2's own page, which cites a post-training set of about 10,000 episodes and 26.3 million pre-training samples.
The training budget is the part that stayed with us. Ai2 says pre-training took about a day on 256 NVIDIA H100 GPUs and fine-tuning about two hours on 64. That is not hobbyist money, but it is within reach of a university lab, which is exactly the audience a fully open release targets.
Then came the follow-up. In May 2026, The Letter Two reported MolmoAct 2, described as 37 times faster than the original and trained on over 30 times as much robot data, including 700+ hours of bimanual demonstrations released openly. The report cites 97.2% on LIBERO for the base model and 98.1% for the Think variant, and says Ai2 is piloting it with Stanford School of Medicine and other partners. Those are still reported results, not independent ones.
Our take: the durable contribution is not the benchmark score, which every release inflates, but the decision to ship data and code alongside weights. Anyone can now test whether reasoning in space beats reasoning in words. The fast sequel suggests Ai2 thinks the answer is yes; we will believe it when someone outside Ai2 reproduces it.
Related articles from LaRebelión
Apple tightens full-disk access to rein in AI agents
OpenAI probes more possible AI hacking incidents after the Hugging Face breach
OpenAI Reports Rogue AI Agents Affecting Over 100 OrganizationsOriginal source: Ai2
Produced with AI support and reviewed by the newsroom