The Allen Institute for AI (Ai2) released MolmoWeb on March 24, an open-source AI agent that can autonomously navigate websites by interpreting screenshots like a human would. Available in 4B and 8B parameter versions, the 8B model scored 78.2% on the WebVoyager benchmark — outperforming agents built on GPT-4o and Claude 3.7 that rely on reading underlying page code.
Unlike proprietary rivals such as OpenAI's ChatGPT Atlas and Perplexity's Comet, MolmoWeb's weights, training data, and code are fully open. Ai2 also released MolmoWebMix, a dataset of 36,000 human task trajectories across 1,100+ websites with 623,000 individual subtask demonstrations — addressing the critical shortage of public training data for web agents.
The release signals a shift in AI agent development toward transparency. By working from screenshots rather than HTML, MolmoWeb is more resilient to webpage changes and uses fewer computational tokens. With test-time scaling, the model reaches 94.7% on WebVoyager, rivaling the best closed systems available today.
