Ai2 Newsletter
June 2026
Top story - Introducing MolmoAct 2—an open foundation for robots that work in the real world
Earlier this month, we released MolmoAct 2, an open model for controlling robots in the real world. Other robotics models have stayed largely closed—some teams behind them release weights, fewer release training data, and almost none publish enough for researchers to study, reproduce, or easily customize the models. MolmoAct 2 changes that, giving the community a foundation it can adapt and improve directly.
Since May 5, MolmoAct 2 and its associated artifacts have been downloaded more than 400K times.
MolmoAct 2 outperforms leading proprietary robotics models on standard benchmarks while running up to 37× faster than its predecessor, MolmoAct. It can complete tasks like moving an apple onto a plate, putting a pipette into a tray, and placing a knife in a box—all without special training for each setup.
Already, MolmoAct 2 has been piloted at the Cong Lab at Stanford School of Medicine, where it handles routine manipulation steps in CRISPR gene-editing experiments—moving samples between stations and operating benchtop equipment. Separately, we hired Cortex AI to independently validate MolmoAct 2's two-armed performance; their evaluation ranked MolmoAct 2 first on 7 of 8 challenging tasks, ahead of four other competing robotics models including Physical Intelligence's π0.5.
MolmoAct 2 ships with the MolmoAct 2-Bimanual YAM dataset, a 700+ hour collection of two-armed robot demonstrations—the largest open-source dataset of its kind ever published, and more than 30× the robot data used in MolmoAct.
Alongside the dataset and model weights, we've released MolmoAct 2's action tokenizer – which translates robot motion into the discrete steps the model predicts – plus training scripts, evaluation rollouts, and a reference hardware setup, making MolmoAct 2 fully open. MolmoAct 2 is also integrated into Hugging Face's LeRobot platform, so teams already in that ecosystem can drop it into their existing setup.
We think open robotics is the future of the field. We're proud to make this strong contribution to it.
NSF OMAI
This month, we brought online the compute infrastructure for the Open Multimodal AI Infrastructure for Science (NSF OMAI) project in partnership with Cirrascale, funded by a $152 million joint investment from NSF and NVIDIA. OMAI will produce foundational open source AI artifacts that others can study and build on.
EMO
EMO is a new modular model whose components naturally specialize during training, organizing themselves around topics like health, news, and politics rather than around surface textual patterns. That means you can load just a small slice of EMO for a specific task and still get near-full performance, making the model more flexible and cheaper to deploy than most.
IFBench
Artificial Analysis integrated our open IFBench evaluation into its Intelligence Index, a cross-model leaderboard widely tracked across the AI industry. IFBench tests how precisely language models follow complex, multi-part instructions, an everyday capability that's often missed by other benchmarks.
AIMIP
The AI Model Intercomparison Project (AIMIP), led by Ai2 with participation from NVIDIA, Google Research, and others, is the first shared benchmark for AI climate models. Phase 1 includes eight model simulations of the global atmosphere and an initial analysis of how they performed.
OlmoEarth v1.1
OlmoEarth v1.1, our latest family of Earth observation models, cut compute costs by up to 3× compared to v1 while keeping similar performance. The savings come from a more efficient way of processing satellite data, making planet-scale tasks like crop-type mapping faster and cheaper for partners.
PointCheck
On Global Accessibility Awareness Day, we featured PointCheck, an independent project that combines Molmo, MolmoWeb, and Olmo 3 to test web accessibility by what users actually see on screen—catching problems that code-level checkers miss. The tool runs entirely on a user's own machine, so no screenshots leave their environment.