Microsoft Fara-7B: New AI Agent for Windows

Microsoft just released Fara-7Bits first small language model designed specifically for using computers as you or I would.

with just 7 billion parametersthis compact model is designed to run directly on your device, meaning less latency, more privacy, and a smoother experience.​

A model that uses your PC like a human

Unlike typical chatbots that only generate text responses, Fara-7B is a computer usage agent (CUA) that actually interacts with your system.

It can click, type, scroll and navigate websites just like you would, but in an automated way.​

The most interesting thing is that you don’t need an entire ecosystem of cloud models to function. While other AI agents require massive servers and multiple subsystems working in the background, Fara-7B is a single, compact and self-sufficient model.

Simply look at a screenshot and makes decisions based on what it sees, without relying on additional information such as accessibility trees or complex analyses.​

How You Were Trained (And Why It Matters)

Microsoft developed a training system called FaraGen that generates synthetic data on a massive scale.

This system makes AI agents perform real tasks in more than 70,000 web domainsimitating human behaviors such as errors, retries, scrolling and searches.​

Each session is reviewed by three separate AI judges to ensure that the steps make sense and that the results match what appears on the screen.

After this rigorous filtering, Microsoft kept 145,630 verified sessions containing more than 1 million individual actions to train the model.​

Performance and efficiency

Here comes the good thing. Fara-7B uses about 124,000 entry tokens and only 1,100 output tokens per task.

Microsoft estimates that completing an entire task costs about 2.5 cents, compared to 30 cents for larger agents. based on GPT-4 or O3.​

As for performance, the numbers are solid for such a light model. Reach a 73.5% in Web Voyager34.1% a OnlineMind 2 Webb, 26.2% a DeepShop y 38.4% a WebTailBench.

This last benchmark is especially relevant because it focuses on real-world tasks like job applications and real estate searches.​

It’s now available (and you can try it)

Fara-7B is available now at Microsoft Foundry and Hugging Face under an MIT license. It also integrates with Magentic-UI, a research prototype from Microsoft Research AI Frontiers.​

But there is more. Microsoft is releasing a quantized, silicon-optimized version specifically for Copilot+ PCs con Windows 11.

This means you can install and test it locally on your computer without depending on the cloud. The pre-optimized package can be downloaded and run directly in community environments.​

By being an open weights model, Microsoft hopes to lower the barriers for developers who want to experiment and advance computer usage agent technology, especially for automate everyday web tasks.​

Related Posts

Leave a Comment