fara

Why we included this project

Teams automating real browser work, not just text responses, get a concrete starting point here: Fara1.5 ships three model sizes (4B, 9B, 27B) that look at screenshots and drive the mouse and keyboard directly, skipping accessibility trees and separate parsing stages. The 9B variant scores 63.4% on Online-Mind2Web and 86.6% on WebVoyager while staying small enough for modest hardware, and the 27B version beats much larger proprietary systems on the same benchmark. Weights are on Hugging Face under an MIT license, so you can self-host, fine-tune, or route through Microsoft Foundry. The repo also includes a CLI for running a model against a live browser, an optional sandboxed setup that logs actions, and the training and evaluation tooling behind the models, useful if you want to reproduce results or score your own agent trajectories.