Ai2 unveils MolmoWeb, an open-weight, fully trained visual browser agent that navigates and acts on a browser by interpreting screenshots, paired with MolmoWebMix, a large dataset for web-based tasks.

MolmoWebMix draws from three data streams: human demonstrations from over a thousand sites with tens of thousands of trajectories and hundreds of thousands of subtasks; synthetic, text-based trajectories including multi-agent task decompositions; and GUI/perception data featuring about 2.2 million screenshot-based QA pairs from roughly 400 websites.

MolmoWeb is positioned against two browser-agent categories—API-only systems and open-weight models—and is claimed to lead on benchmarks like WebVoyager, Online-Mind2Web, DeepShop, and WebTailBench, surpassing older GPT-4o-based agents that rely on accessibility trees.

The approach emphasizes visual grounding, arguing that screenshots offer a compact, stable representation that is easier to interpret and debug than traditional page representations.

Ai2 notes enterprise considerations such as auditability, fine-tuning capability, and avoiding per-call API dependencies when choosing browser agents.

Limitations include OCR/read errors on screenshots, unreliable drag-and-drop interactions, and weaker performance on ambiguous prompts; the model has not been trained on logins or financial transactions.

Additional limitations involve occasional hallucinations from reading text in screenshots, potential misdirection from wrong actions, and reduced performance under ambiguous constraints or tasks involving PII, logins, or financial transactions.

MolmoWeb is browser-agnostic, capable of running on local Chrome or Safari or hosted services, with a cloud demo powered by Browserbase.

The rollout occurs amid an industry push for AI agents that can navigate computers and the web, with notable leadership changes at Ai2 and ongoing 2026 funding commitments.

MolmoWeb is designed to perform a wide range of web tasks without APIs and supports user-developed extensions, while cautioning against handling sensitive tasks like logins or financial transactions in self-hosted deployments.

The dataset enables end-to-end training by recording trajectories so the agent can learn to perform tasks from visual input.

Openly released as part of Ai2’s broader open-Language Model initiative, MolmoWeb emphasizes openness to foster community development.