Introduced by Microsoft FARA-7B, a new 7 billion parameter model Designed to act as a computer use agent (CUA) capable of performing complex tasks directly on the user’s device. FARA-7B sets state-of-the-art results for its size, providing a way to build AI agents that don’t rely on large-scale, cloud-dependent models and can run on compact systems with low latency and better privacy.
Although the model is an experimental release, its architecture addresses a key barrier to enterprise adoption: data security. Because FARA-7B is small enough to run locally, it allows users to automate sensitive workflows, such as managing internal accounts or processing sensitive company data, without ever having that information exposed.
How FARA-7B sees the web
FARA-7B is designed to navigate the user interface using a single set of tools: a mouse and keyboard. The model works by visually understanding a web page through screenshots and predicting specific coordinates for actions such as clicking, typing and scrolling.
Importantly, FARA-7B does not depend on "access trees,” the underlying code structure that browsers use to describe Web pages to readers. Instead, it relies entirely on pixel-level visual data. This approach allows agents to interact with Web sites even when the underlying code is inconsistent or complex.
According to Yash Lara, a senior PM at Microsoft Research, processing processes all visual input. "pixel autonomy," Since the logic required for screenshots and automation remains on the user’s device. "This approach helps organizations meet stringent requirements in regulatory areas, including HIPAA and GLBA." He told VentureBeat in written comments.
In benchmarking tests, this visual-first approach has yielded strong results. But Web browsera benchmark for web agents, FARA-7B achieved a task success rate of 73.5%. It optimizes large, resource-intensive systems, including GPT-4Owhen indicated to act as a computer use agent (65.1%) and local UI-TARS-1.5-7B model (66.4%).
Versatility is another key differentiator. In comparative tests, the FARA-7B completed tasks in an average of about 16 steps compared to about 41 steps for the UI-TARS-1.5-7B model.
Handling hazards
However, the transition to autonomous agents is not without risks. Microsoft notes that FARA7B shares limitations common to other AI models, including possible hallucinations, errors in following complex instructions, and degradation of accuracy on complex tasks.
To mitigate these risks, the model was trained to recognize "Critical points" A critical point is defined as any situation that requires a user’s personal data or consent before an irreversible action takes place, such as sending an email or completing a financial transaction. Upon reaching such a juncture, FARA-7B is designed to stop and expressly request user approval before proceeding.
Managing this interaction without frustrating the user is a key design challenge. "Balancing strong security measures with a smooth user journey is key," Lara said. "Having a UI, like Microsoft Research’s Mechanical UI, is critical to providing opportunities for users to intervene when needed, while also helping to avoid approval fatigue." Magic-U is a research prototype specifically designed to facilitate these human agent interactions. FARA-7B is designed to run in the Magick UI.
Eliminating complexity in a single model
An increasing trend in the development of FARA-7B is highlighted Ease of knowledgewhere the capabilities of a complex system are compressed into a smaller, more efficient model.
Building a CUA usually requires a large amount of training data that represents how the web is navigated. Collecting this data through human interpretation is prohibitively expensive. To solve this, Microsoft used a synthetic data pipeline built on The magical onea multi-agent framework. In this setup, a "The orchestrator" The agent plans and directs "Web Surfer" Agents generate 145,000 successful task trajectories to browse the web.
After that the researchers "Ast" This complex interaction data in FARA-7B, built on QWEN2.5-VL-7B, is a base model chosen for its long context window (up to 128,000 tokens) and its strong ability to integrate textual instructions with visual elements on the screen. Although data generation requires a heavy multi-agent system, FARA-7B is a single model by itself, demonstrating that a small model can effectively learn sophisticated behaviors at runtime without the need for complex scaffolding.
The training process relied on supervised fine-tuning, where the model learns by simulating successful examples generated by a synthetic pipeline.
are waiting
Although the current version was trained on static datasets, future iterations will focus on making the model better, not necessarily larger. "Going forward, we will try to keep our models small," Lara said. "Our ongoing research is focused on making models of agents better and safer, not more." This includes looking for techniques Reinforcement learning (RL) in a live, sandboxed environment, which allows the model to be learned by trial and error in real time.
Microsoft has made the model available on Hug Face and Microsoft Foundry under the MIT license. However, Lara cautions that while the license allows for commercial use, the model is not yet ready for production. "You may freely experiment and prototype with FARA-7B under the MIT License," he says, "But it is best suited for pilots and proofs rather than mission-critical deployments."