What is Fara-7B
Fara-7B is Microsoft's open-source small language model (SLM) focused on computer use as an agent. The model perceives web pages visually and operates interface elements such as the mouse and keyboard to help users complete tasks, such as filling out forms, searching for information, or booking trips. With only 7 billion parameters, the model achieves top performance in its size class and can run directly on-device, reducing latency and enhancing privacy protection. Fara-7B is trained on synthetic data, has efficient task execution capabilities, and also emphasizes safety. The model is currently released as an experimental version, aiming to invite the community to explore and provide feedback to advance technological progress.
Main Features of Fara-7B
-
Automated Task Execution: By operating interface elements such as the mouse and keyboard, it helps users complete various computer tasks, such as filling out forms, searching for information, booking travel, managing accounts, and more.
-
Visual Perception and Interaction: The model can directly perceive web page content and perform operations such as clicking, scrolling, and typing by predicting coordinates, relying entirely on screen visual information without additional auxiliary information (such as accessibility trees).
-
User Interaction and Control: When executing tasks, Fara-7B pauses at key points (such as when user privacy or important decisions are involved) and waits for user confirmation or input, ensuring the user has complete control over the task.
-
Privacy Protection and Security: All operations are completed on the local device, user data is not transmitted externally, and logging and sandboxed environments ensure privacy and security.
-
Efficient Task Execution: Through an optimized model architecture and training method, Fara-7B can complete tasks in fewer steps, making it more efficient and lower-cost than other similar models.
Technical Principles of Fara-7B
-
Vision-Based Interaction: The model perceives web page content through screenshots and directly simulates the way human users interact with computers, without relying on the web page's internal structure (such as the DOM tree or accessibility tree).
-
Synthetic Data Generation: To address the lack of high-quality training data, Microsoft developed a synthetic data generation pipeline. The pipeline generates large-scale multi-step task data from public web pages and task prompts, avoiding the high cost of manual annotation.
-
Multi-Agent System Training: During the training phase, Fara-7B uses a multi-agent system to solve synthetic tasks and generate demonstration data for supervised fine-tuning. The agents include a task planner, a web operator, and a user simulator, among others, which complete tasks through collaborative work.
-
Single-Model Distillation: The complexity of the multi-agent system is distilled into a single model, enabling Fara-7B to run independently and simplifying deployment and use.
-
Enhanced Safety Mechanisms: Safety data is incorporated into the model's training, enabling it to refuse harmful tasks. At the same time, all operations are logged and auditable, ensuring users have complete control over the model's behavior.
Fara-7B Project Address
- Project Website：https://www.microsoft.com/en-us/research/blog/fara-7b-an-efficient-agentic-model-for-computer-use/
- GitHub Repository：https://github.com/microsoft/fara
- HuggingFace Model Library：https://huggingface.co/microsoft/Fara-7B
- Technical Paper：https://www.microsoft.com/en-us/research/wp-content/uploads/2025/11/Fara-7B-An-Efficient-Agentic-Model-for-Computer-Use.pdf
Application Scenarios of Fara-7B
-
Office Automation: Fara-7B can automatically process documents, write emails, and enter data, significantly improving office efficiency.
-
Information Retrieval and Organization: The model can quickly search for online information and organize and summarize it, helping users efficiently obtain the materials they need.
-
E-commerce: Fara-7B can automatically search for products, compare prices, and assist with placing orders, optimizing the shopping experience.
-
Travel planning: The model can intelligently arrange itineraries and book flights and hotels, simplifying the travel preparation process.
-
Online learning: Fara-7B can automatically search for courses and organize learning materials, helping users learn efficiently.