You have spent years watching billion-dollar companies build AI tools that require their servers, their APIs, their pricing plans, and their permission to use.
On March 24, 2026, a nonprofit research institute released a web agent that navigates browsers, fills out forms, searches flights, and completes complex tasks online — with no API fees, no closed access, no gatekeepers — and it outperformed Anthropic's Claude 3.7 and OpenAI's CUA on benchmarks.
The future of AI does not belong only to the companies with the biggest checkbooks. It belongs to anyone willing to build it.
The Allen Institute for AI, known as Ai2, released MolmoWeb on March 24, 2026 — a fully open-source visual web agent built on two compact models: MolmoWeb-4B and MolmoWeb-8B, according to the official Ai2 blog. The models are available on Hugging Face at allenai/MolmoWeb-4B and allenai/MolmoWeb-8B, and the full code is published at github.com/allenai/molmoweb under the Apache 2.0 license — meaning anyone can download, modify, deploy, and build on it, for free, forever. The 8B model scored 78.2% on WebVoyager, approaching OpenAI's proprietary o3 model at 79.3%, according to The Decoder. On the ScreenSpot visual grounding benchmarks, the MolmoWeb-8B grounding model outperformed Claude 3.7 and OpenAI CUA — both closed, expensive, proprietary systems — according to the official Ai2 research paper. This is not an incremental research update. This is a declaration that the open-source community can now build autonomous AI agents that compete at the frontier — on their own hardware, at their own pace, answering to no one.
What Is MolmoWeb — and Who Built It?
MolmoWeb is an open-source visual web agent that controls a browser the same way a human does: by looking at screenshots. It receives a task instruction — such as "Find the cheapest nonstop flights from Seattle to Tokyo" — a screenshot of the current browser state, and the history of recent actions. From that, it produces a brief chain-of-thought reasoning step and then executes a browser action: clicking, typing, scrolling, or navigating. No HTML parsing. No DOM access. No proprietary APIs. Just pixels, reasoning, and action.
The agent was built by the Allen Institute for AI, a Seattle-based nonprofit research organization founded in 2014 by Paul Allen, co-founder of Microsoft. Ai2's mission, as stated by interim CEO Peter Clark, is "advancing AI research and engineering for the common good, and turning our open breakthroughs into lasting, real-world impact." The MolmoWeb lead researcher is Tanmay Gupta, Ai2's senior research scientist, who led the project alongside the broader Molmo team.
MolmoWeb is built on Molmo 2, Ai2's most recent open multimodal model — which uses a Qwen3 language model and a SigLIP2 vision encoder. It was trained on 64 H100 GPUs for up to 50,000 steps. The training data comes entirely from open sources — no distillation from proprietary models like GPT-4 or Claude was used, according to the official Ai2 paper.
What Happened: March 24, 2026
On March 24, 2026, Ai2 released MolmoWeb in two sizes alongside the complete open-source stack: model weights, training data (MolmoWebMix), evaluation tools, inference code, and a live demo at molmoweb.allen.ai. Training code was announced as coming soon.
MolmoWebMix — the training dataset released alongside the model — is the largest publicly released dataset of human web task execution ever assembled. It includes 30,000 human-completed web tasks across more than 1,100 websites, representing over 590,000 individual browser actions recorded via a custom Chrome extension, according to Ai2. The dataset also includes more than 2.2 million screenshot-question-answer pairs for visual grounding, and synthetic trajectories generated by a three-model pipeline: a Gemini 2.5 Flash planner, an operator agent, and a GPT-4o verifier, according to The Decoder.
The timing was striking. MolmoWeb launched the same day Microsoft confirmed it had hired Ai2's recently departed CEO, Ali Farhadi, and several key researchers — including Hanna Hajishirzi, who led the OLMo language model project — for Mustafa Suleyman's MAI Superintelligence team, according to Aihola. The release was Ai2's answer: even as key talent moved to big tech, the open-source community gets the work.
The Numbers That Matter
- 78.2%: MolmoWeb-8B score on WebVoyager (vs. 79.3% for OpenAI o3 — a near tie)
- 94.7%: MolmoWeb-8B score on WebVoyager with pass@4 test-time scaling
- 60.5%: MolmoWeb-8B on Online-Mind2Web with pass@4 (vs. 35.3% with single rollout)
- 42.3%: MolmoWeb-8B on DeepShop — within 6 points of GPT-5, according to The Decoder
- 49.5%: MolmoWeb-8B on WebTailBench
- Top-ranked: ScreenSpot and ScreenSpot v2 visual grounding — beats Claude 3.7 and OpenAI CUA
- 4B and 8B: Parameter sizes — small enough to run locally on consumer hardware
- 30,000: Human web task demonstrations in MolmoWebMix — the largest public dataset of its kind
- 590,000+: Individual browser actions in the dataset
- 1,100+: Websites represented in MolmoWebMix
- 2.2 million: Screenshot-QA pairs for visual grounding training
- Apache 2.0: License — fully free for commercial and research use
| Benchmark | MolmoWeb-8B | Best Proprietary Comparison | |---|---|---| | WebVoyager | 78.2% (94.7% pass@4) | OpenAI o3: 79.3% | | Online-Mind2Web | 35.3% (60.5% pass@4) | GPT-5 class agents | | DeepShop | 42.3% | GPT-5: ~48% | | ScreenSpot v2 (grounding) | Best open model | Beats Claude 3.7 + OpenAI CUA |
Why This Matters to You
You have worked hard to build your skills, your business, your technical knowledge. You know how to use AI tools — you have put in the hours learning what others haven't bothered to understand yet. That effort gives you something that no subscription can buy: the foundation to actually benefit from what MolmoWeb represents.
Here is what it means in practice. Every enterprise-grade web agent today — OpenAI Operator, Google Project Mariner, Anthropic's Computer Use — is locked behind API pricing, closed terms, and controlled access. You can use them, but you cannot inspect them, modify them, or own them. MolmoWeb changes that equation completely. You can download the weights right now. You can run it on your own machine, point it at any website, and let it work. You can modify it for your specific workflow. You can build a product on top of it without paying per-call to any company.
For developers and entrepreneurs: Tanmay Gupta, Ai2's senior research scientist, suggested specific use cases — automating repetitive browser workflows "that need to be run at a predictable frequency," such as pulling data from websites weekly, or scaling to hundreds of parallel queries like gathering author metrics across large research databases, according to The AI Economy. For businesses that currently pay human workers or expensive API subscriptions to perform repetitive web tasks, MolmoWeb represents a direct path to automation that you own and control.
The pride in building something with MolmoWeb is different from clicking "generate" in a chatbot. You are choosing to build on an open foundation — the same way the internet was built, the same way Linux was built. That is the kind of work that compounds.
What Comes Next
Ai2 confirmed that all 2026 programs remain fully funded, including MolmoWeb's continued development. Training code — not yet released at launch — is expected to follow. The MolmoWebMix dataset pipeline is fully described in the research paper (arXiv 2601.10611), enabling the community to generate new training data independently.
The research paper identified two clear paths to further performance improvement: increasing the number of parallel rollouts (pass@4 already reaches 94.7% on WebVoyager) and extending the number of inference steps per task. Both are accessible to anyone running the open-weight models — no proprietary infrastructure required.
Ai2 also maintains three other open agents alongside MolmoWeb: DR Tolu for long-form deep research, Sera for coding, and Asta for scientific research — forming a growing open-source ecosystem of specialized AI agents, according to The AI Economy. The community that builds on MolmoWeb today is the one that will define what autonomous AI agents look like for everyone else tomorrow.
Want to follow every major AI move the moment it happens? Join thousands of readers at fabioraphaelli.com and get the biggest AI news every week.
Frequently Asked Questions
What is MolmoWeb and who created it?
MolmoWeb is a fully open-source visual web agent released on March 24, 2026, by the Allen Institute for AI (Ai2), a nonprofit research organization based in Seattle, founded in 2014 by Paul Allen. It comes in two sizes — 4B and 8B parameters — and controls a web browser by interpreting screenshots, without needing access to HTML, page structure, or proprietary APIs. All model weights, training data, code, and evaluation tools are available under the Apache 2.0 license at github.com/allenai/molmoweb and on Hugging Face.
How does MolmoWeb compare to OpenAI and Anthropic's web agents?
MolmoWeb-8B scored 78.2% on WebVoyager, approaching OpenAI's proprietary o3 model at 79.3%, according to The Decoder. On the ScreenSpot and ScreenSpot v2 visual grounding benchmarks, a dedicated MolmoWeb-8B grounding model outperformed Claude 3.7 Computer Use and OpenAI CUA, according to the official Ai2 research paper. MolmoWeb also outperformed agents built on the larger GPT-4o that had access to annotated screenshots and structured page data — which Ai2 called "a striking result given that those models enjoy orders-of-magnitude higher parameters."
Can I run MolmoWeb on my own computer?
Yes. MolmoWeb was designed for self-hosted deployment — either locally or on cloud services. Both the 4B and 8B models are small enough to run on accessible hardware. The inference package includes a high-level Python client that manages browser sessions end-to-end. You need Python 3.10–3.12, the uv package manager, and Playwright for browser control. Detailed installation instructions are available at github.com/allenai/molmoweb. Ai2 recommends avoiding tasks that require sensitive personal information like passwords or financial credentials.
What is MolmoWebMix?
MolmoWebMix is the training dataset released alongside MolmoWeb on March 24, 2026. It is the largest publicly released dataset of human web task execution, containing 30,000 human-completed web tasks across more than 1,100 websites, over 590,000 individual browser actions, and more than 2.2 million screenshot-question-answer pairs for visual grounding training. It also includes synthetic trajectories generated by an automated three-model pipeline using Gemini 2.5 Flash, an operator agent, and a GPT-4o verifier, according to The Decoder. The full dataset is available open-source alongside the model.
What tasks can MolmoWeb currently perform?
MolmoWeb can perform a wide range of web tasks that do not require sensitive personal credentials. Demonstrated capabilities include navigating websites like GitHub, Google Flights, and Amazon; searching for products and comparing prices; completing multi-step research tasks; filling out forms; scrolling and clicking through complex interfaces; and following multi-step task instructions across different websites. Senior research scientist Tanmay Gupta specifically cited automating repetitive data-collection tasks at scale — such as querying hundreds of author profiles — as a practical enterprise use case, according to The AI Economy.









