microsoft/Fara-7B

Fara-7B: An Efficient Agentic Model for Computer Use

Fara-7B, developed by Microsoft Research, is a 7 billion parameter multimodal decoder-only language model optimized for computer use as an agent. It functions as an ultra-compact Computer Use Agent (CUA), taking user goals (text) and current screenshots as input to directly predict thoughts and actions with grounded arguments. This model is particularly adept at automating web tasks, planning, and executing high-level goals such as booking restaurants, applying for jobs, or managing shopping lists.

Key Capabilities

• Web Task Automation: Specializes in automating interactions on websites, including shopping, travel booking, restaurant reservations, and information seeking.
• Multimodal Understanding: Processes browser screenshots alongside text context to perceive web page layouts and execute actions.
• Agentic Planning & Execution: Plans and executes complex, high-level tasks by predicting sequential actions and their arguments (e.g., coordinates for clicks).
• Efficiency: Achieves competitive performance with larger agentic systems despite its compact 7 billion parameter size.
• Privacy: Designed for on-device execution, offering enhanced privacy and lower latency for automated tasks.

Good For

• Developers building applications that require automated web interaction and task completion.
• Use cases involving step-by-step action execution based on visual and textual cues from web interfaces.
• Scenarios where an efficient, compact agentic model is preferred for on-device deployment or resource-constrained environments.
• Research and development in agentic AI, particularly for computer control and web automation.

Fara-7B is built upon Qwen 2.5-VL (7B) and was trained on a large-scale, fully synthetic dataset of action trajectories. It incorporates robust safety measures, including critical point recognition to halt actions requiring user permission or sensitive information, and is trained to refuse harmful tasks.