Krux
March 26, 2026
Open-Source MolmoWeb Rivals GPT-4o on Browser Automation
Published: March 26, 2026 at 12:34 AM
Updated: March 26, 2026 at 12:34 AM
100-word summary
Allen Institute just released MolmoWeb, an open-source web agent that reads screenshots and autonomously clicks, types, and scrolls through websites to complete tasks. The 8B-parameter version reportedly matches GPT-4o on web-navigation benchmarks, but you can actually inspect the code and run it yourself. Unlike closed API services, MolmoWeb runs locally using Playwright for browser control. That means you can automate form submissions, scrape competitor sites, or run end-to-end tests without vendor lock-in. The tradeoff: you'll need to clone a GitHub repo, wrangle dependencies, and spin up your own server. The timing matters. As OpenAI and Google push their own browser agents, having an auditable alternative shifts who gets to build...
What happened
Allen Institute just released MolmoWeb, an open-source web agent that reads screenshots and autonomously clicks, types, and scrolls through websites to complete tasks. The 8B-parameter version reportedly matches GPT-4o on web-navigation benchmarks, but you can actually inspect the code and run it yourself. Unlike closed API services, MolmoWeb runs locally using Playwright for browser control. That means you can automate form submissions, scrape competitor sites, or run end-to-end tests without vendor lock-in. The tradeoff: you'll need to clone a GitHub repo, wrangle dependencies, and spin up your own server.
Why it matters
The timing matters. As OpenAI and Google push their own browser agents, having an auditable alternative shifts who gets to build and verify web automation.