Trending Now - Ai2 Unveils Molmoweb, a New Open-source Visual Web Agent

WHO:

Ai2, a nonprofit organization based in Seattle, is responsible for the release of MolmoWeb.

Tanmay Gupta, senior research scientist at Ai2, provided insights about MolmoWeb.

WHAT:

Ai2 launched MolmoWeb, an open-source visual web agent available in 4 billion and 8 billion parameter sizes.

MolmoWeb operates by interpreting browser screenshots rather than parsing HTML.

The accompanying dataset, MolmoWebMix, includes 30,000 human task trajectories, 590,000 individual subtask demonstrations, and 2.2 million screenshot question-answer pairs.

MolmoWeb is described as the largest publicly released collection of human web-task execution ever assembled.

MolmoWeb outperforms older API-based agents built on GPT-4o with accessibility tree plus screenshot input.

WHERE:

The release is associated with the Seattle-based nonprofit Ai2.

WHEN:

MolmoWeb was released in March 2026.

WHY:

The release aims to provide a transparent alternative to closed systems from OpenAI, Google, and Anthropic.

Ai2 seeks to offer a fully trained open-weight vision model that allows for auditing and fine-tuning on internal workflows.

HOW:

MolmoWeb operates by receiving a task instruction, the current screenshot, a text log of previous actions, and the current URL and page title to produce a natural-language thought describing its reasoning.

The model executes actions such as clicking, typing, scrolling, navigating, or switching tabs based on the received inputs.

The training dataset includes human demonstrations recorded via a custom Chrome extension, synthetic trajectories generated using text-based accessibility-tree agents, and GUI perception data for reasoning about page content from images.

(UPDATED March 24, 2026 17:54 UTC)

Today, Ai2, a nonprofit organization based in Seattle, launched MolmoWeb, an open-source visual web agent that comes in both four billion and eight billion parameter sizes. This innovative tool interprets browser screenshots instead of parsing HTML, and it boasts the largest publicly released collection of human web-task execution data, including 30,000 human task trajectories and over two million question-answer pairs. The release aims to provide a transparent alternative to existing closed systems, allowing for auditing and fine-tuning on internal workflows, while outperforming older API-based agents built on GPT-4o.