Ai2 Releases MolmoWeb: Open-Source Visual Web Agent with 30K Human Trajectories

Allen Institute for AI·medium signal

The Allen Institute for AI released MolmoWeb, an open-source visual web agent in 4B and 8B parameter sizes built on Molmo 2. It navigates browsers by interpreting screenshots (not DOM/HTML), handling clicks, typing, and scrolling. The release includes MolmoWebMix — 30,000 human task trajectories across 1,100+ websites, 590,000 subtask demonstrations, and 2.2M screenshot Q&A pairs. MolmoWeb outperforms all open models on web agent benchmarks and approaches proprietary systems from OpenAI, with full weights, training data, and evaluation tools released openly.