Microsoft Releases Fara1.5-27B: A Vision-Only Multimodal Browser Automation Agent
Microsoft Research has released Fara1.5-27B, a vision-only multimodal computer use agent fine-tuned from Qwen3.5-27B designed to automate web browser tasks. The model operates by analyzing screenshots and emitting structured tool calls, such as clicks and scrolls, to complete tasks end-to-end. This release advances the field of computer use agents by demonstrating that effective browser automation can be achieved purely through visual perception without relying on the DOM or accessibility trees. It provides a specialized, open-weights option for developers looking to build local, efficient web-automation workflows. Fara1.5-27B was trained on synthetic data generated by the FaraGen1.5 pipeline and is co-designed to run with MagenticLite, a lightweight agentic application. However, it is limited to English, susceptible to visual prompt injections, and prone to error accumulation over multi-step trajectories.
## BACKGROUND
Computer use agents (CUAs) are AI systems designed to interact with digital interfaces in a human-like manner, such as controlling a cursor or typing. While many traditional web agents rely on parsing the Document Object Model (DOM) or accessibility trees of a webpage, vision-only agents interact with the interface solely by processing visual pixels, making them more adaptable to different layouts but also vulnerable to visual rendering errors.