Browser-Use Trains BU-30B-A3B-Preview to Run 200 Web Tasks per Dollar
Browser-Use ships its own open-source browser agent model and a public benchmark, staking a position at the center of the web automation stack
Takeaways
- Browser-Use released BU-30B-A3B-Preview, their first open-source LLM fine-tuned for browser automation, hostable on a single GPU.
- The model uses a Mixture-of-Experts architecture: 30B total parameters, only 3B active at inference — delivering 200 tasks per dollar.
- Browser-Use launched BU Bench V1, an open benchmark of 100 real-world browser tasks across live websites.
- Claude Fable 5 tops the open-source leaderboard at 80%; Browser-Use Cloud ranks #1 on Odysseys with 87.4% on 200 long-horizon tasks.
- The
ChatBrowserUseclient now routes to any provider (OpenAI, Anthropic, Google) with a single API key, plus their own bu-* models. - The repo has crossed 83,000 stars with an MIT license; cloud tier is free to start.
Browser-Use, the MIT-licensed Python framework for connecting LLMs to real browsers, has grown into something bigger than a library. With over 83,000 GitHub stars and nearly 10,000 forks, the project built by Magnus and Gregor , two developers based in Zurich and San Francisco , has released its own fine-tuned model, an open benchmark, a CLI, and a hosted cloud tier. This is no longer just a Playwright wrapper. It is a full-stack web agent platform.
The model they trained themselves
The most significant development is BU-30B-A3B-Preview, Browser-Use's first self-trained open-source LLM. It uses a Mixture-of-Experts (MoE) architecture , a design where the model has 30B total parameters but only activates 3B of them at inference time. MoE lets you get the reasoning quality of a large model at the compute cost of a small one. The result is a model that brings state-of-the-art browser automation capabilities while fitting on a single GPU.
The model is fine-tuned from Alibaba Cloud's Qwen3-VL-30B-A3B-Instruct, optimized specifically for browser automation scenarios, with multimodal input support. Its DOM understanding and visual reasoning capabilities let AI agents parse page layouts and screenshots the way a human would, enabling reliable automation execution. The cost story is compelling: at 30B parameters with 3B active, it delivers 200 tasks per dollar , 4x more cost-efficient than BU 1.0.
You can run it self-hosted via vLLM, or access it through the Browser-Use Cloud API under the same