Ai2 releases open-source web agent MolmoWeb to challenge OpenAI, Google, and Anthropic

Ai2 Launches MolmoWeb as Open-Source Web Agent Alternative

The Allen Institute for AI (Ai2) has released MolmoWeb, an open-source web agent designed to rival closed systems from OpenAI, Google, and Anthropic1

1

. This visual AI agent can take control of web browsers and automate tasks by interpreting screenshots of webpages the way a person would, rather than relying on underlying page code2

2

. Built on Ai2's Molmo 2 multimodal model family, MolmoWeb arrives at a critical moment when major tech companies are racing to build AI agents capable of navigating computers and the web on behalf of users1

1

.

The Seattle-based nonprofit institute released MolmoWeb in two sizes: 4B and 8B parameters, making it available for free along with weights, training data, code, and evaluation tools2

2

. Developers can access the agent through Hugging Face and GitHub, along with a demo for testing on supported websites1

1

. This community-driven alternative lets researchers and developers look under the hood to understand what's happening in ways not possible with closed systems1

1

.

How MolmoWeb Works to Automate Tasks

MolmoWeb operates by observing web pages through a series of screenshots and then interacting directly via the interface by predicting what will happen when it takes actions such as clicking, typing characters into text fields, or scrolling up and down2

2

. The web agent supports navigating URLs, clicking on screen coordinates, typing text into fields, scrolling through pages, opening and switching browser tabs, and sending messages back to users2

2

.

This visual approach offers distinct advantages. Ai2 designed the agent this way so it won't break if the underlying webpage code or HTML changes on the fly, as some web pages obfuscate how they operate to protect themselves or use specialized JavaScript engines to detect bots and stop ad blockers2

2

. Using underlying code can also consume tens of thousands of tokens, the essential currency of AI operations2

2

. Visual interfaces behave more closely to how humans interact with web interfaces, making it easier to debug why the model did what it did2

2

.

MolmoWeb Outperforms GPT-4o on Benchmarks

Despite its compact size, MolmoWeb achieves state-of-the-art results among open-weight web agents. When tested on popular evaluation suites, the 8B model scored 78.2% on WebVoyager, 42.3% on DeepShop, and 49.5% on TailBench, outperforming leading open-weight models such as Fara-7B across all four benchmarks2

2

. Ai2 reports that the 8B version outperformed agents built on much larger proprietary models including GPT-4o on key web navigation tasks1

1

.

Ai2 highlighted that MolmoWeb can outperform agents built on GPT-4 that rely on annotated and structured page data, a particularly important result given that those models can "see" deeply into the very code of the webpage and have substantially larger parameter sizes2

2

. Unlike other open-weight web agents, MolmoWeb was trained without compressing a proprietary vision-based agent, with data coming from synthetically generated text-only accessibility agents and human usage of actual web browsing activities2

2

.

Building on Olmo's Open Foundation for LLMs

"In many ways, web agents today are where LLMs were before Olmo -- the community needs an open foundation to build on," Ai2 stated in a blog post, referring to its open large language model project that has served as a counterpoint to closed models from OpenAI and others1

1

. This release provides more access to open-weight browser AI agents, helping researchers and hobbyists develop their own web automations2

2

.

The release comes during a transition period for Ai2, with CEO Ali Farhadi and key researchers departing for Microsoft, where they are joining Mustafa Suleyman's Superintelligence team1

1

. Ai2's primary funder is shifting its focus away from model training toward real-world applications of AI, though all of Ai2's programs for 2026 are fully funded1

1

. Anthropic recently acquired Seattle-based startup Vercept, founded by Ai2 veterans, which was building similar screen-understanding agentic technology for Macs and PCs1

1

. Closed-source providers including Perplexity AI have already entered the market with agentic web browsers capable of automating web tasks2

2

.