MolmoWeb

MolmoWeb is an open-source visual web agent from Ai2 that automates browser-based tasks by interpreting screenshots alone, without relying on DOM or accessibility tree data. Built on the Molmo 2 multimodal model family, it comes in 4B and 8B parameter sizes. The agent operates in a simple loop: given a task instruction and a live webpage screenshot, it predicts the next action—click, type, scroll, navigate, open/switch tabs—and executes it, all while keeping its reasoning trace visible to the user. Designed for self-hosted deployment on local or cloud servers, it targets developers and researchers who need a transparent, reproducible alternative to proprietary web agents like Operator or Project Mariner. MolmoWeb's training dataset, MolmoWebMix, is the largest public web agent training dataset, combining 36K human task trajectories across 1,100+ websites with synthetic trajectories from text-only accessibility-tree agents. Notably, no distillation from proprietary vision-based agents was used—all training data is open. The full codebase released on April 10 includes training code, an evaluation harness, an annotation tool for human demonstrations, and a synthetic data generation pipeline. Evaluation harness supports benchmarks such as WebVoyager, Online-Mind2Web, WebTailBench, and Deepshop. Key features include operation via screenshots only (no DOM), two model sizes, support for multiple browser actions, self-hosted deployment, full codebase release, and the MolmoWebMix dataset. The annotation tool allows users to record their own human demonstrations and fine-tune the model on domain-specific tasks. The demo UI code is provided as a starting point for custom interfaces. Compared to proprietary agents like OpenAI's Operator or Google's Project Mariner, MolmoWeb offers complete transparency but requires technical expertise to deploy. It fills the same gap for web agents that Olmo did for LLMs—an open foundation for reproducible research.