Ai2 releases MolmoWeb, an open-weight visual web agent with 30K human task trajectories and a full training stack

Ai2 released MolmoWeb as an open-weight visual web agent, ai2 released MolmoWeb in 4B and 8B parameter sizes.

MolmoWebMix includes 30,000 human task trajectories, molmoWebMix contains data from more than 1,100 websites.

Contains 590,000 individual subtask demonstrations, molmoWebMix includes recordings of subtasks.

Includes 2.2 million screenshot question-answer pairs, gUI perception component covers element grounding and screenshot reasoning.

MolmoWeb processes task instruction, screenshot, action log, URL and title to produce a reasoning step and execute a browser action, molmoWeb operates from browser screenshots without parsing HTML or accessibility trees.