Microsoft Fara-7B: The AI Model That Can Use Your Computer Like a Human

Artificial intelligence has spent the last decade learning to read, write, and generate images. The next frontier is something far more practical — teaching AI to actually use a computer the way a human does: by looking at the screen, deciding what to do next, and clicking the right buttons in the right order. Microsoft Research has taken a serious step in that direction with Fara-7B, a 7-billion parameter agentic small language model purpose-built for computer use.

What makes Fara-7B different from every other AI model you have heard about is its fundamental approach. Most AI systems interact with the digital world through structured APIs, JSON responses, or text-based interfaces. Fara-7B throws all of that away. It perceives the computer exclusively through raw screenshots — just pixels on a screen — and then predicts precise mouse clicks, keyboard inputs, and scrolling actions to accomplish multi-step tasks on real websites. It is, in the simplest terms, an AI that navigates the internet the same way you do, but it does it at machine speed and it never gets tired.

What Exactly Is Fara-7B

Fara-7B stands for Functional Agentic Reasoning Agent with 7 billion parameters. It was released by Microsoft Research on November 24, 2025, as an open-weight model under the MIT license, which means anyone can download it, use it commercially, and modify it without restrictions. The model is available on both HuggingFace and Microsoft Foundry, making it accessible to individual researchers, startups, and large enterprises alike.

At its core, Fara-7B is built on top of the Qwen2.5-VL-7B vision-language model from Alibaba. Microsoft took this strong multimodal foundation and trained it specifically for agentic computer use — the ability to look at a web page, understand its layout, reason about what needs to happen next, and then execute the required actions through coordinate-based mouse clicks and keystrokes. The result is a model that does not rely on accessibility trees, DOM structures, or any other structured representation of the web. It sees what you see, and it acts the way you act.

This is a critical distinction. Previous approaches to web agents either required access to the underlying HTML structure of a page, which made them fragile and dependent on website architecture, or they were massive frontier models that required enormous compute resources. Fara-7B occupies a completely different position: it is small enough to run on a single consumer GPU or even on-device, yet it delivers performance that rivals or exceeds models many times its size.

How Fara-7B Works Under the Hood

The architecture of Fara-7B follows what Microsoft calls the observe-think-act loop, a continuous cycle that mirrors how a human interacts with a computer. In every cycle, the model receives a task instruction along with a history of previous actions and a set of the most recent screenshots — typically the last three. It then reasons about what it sees on screen and decides the single next action to take.

The perception component of Fara-7B is entirely pixel-based. Unlike earlier systems that parsed accessibility trees or HTML elements to identify clickable buttons and input fields, Fara-7B looks at the raw screenshot and identifies interactive elements through visual understanding. It sees a search bar as a rectangular text input field, a submit button as a colored rectangle with text, and a navigation menu as a row of clickable text labels. The model predicts the exact pixel coordinates where each action should occur.

For actions, Fara-7B uses a set of atomic operations provided through a tool interface built on the Playwright browser automation framework. These actions include mouse clicks at specific coordinates, keyboard typing into focused elements, scrolling to navigate long pages, and browser-level macros like performing a web search or navigating to a URL. The model outputs a single action per step, and the environment executes it, producing a new screenshot that feeds back into the next cycle.

This approach has several important advantages. Because the model works with raw pixels, it can handle any website regardless of its underlying technology — whether it is built with React, Angular, plain HTML, or even Flash. It does not matter how the page is structured internally because the model reads it the same way a human does: by looking at it. This makes Fara-7B fundamentally more robust than DOM-based agents, which break whenever a website changes its HTML structure.

The task memory of Fara-7B is another critical component. Each interaction cycle receives a full history that includes all user-provided instructions, the complete sequence of previous actions taken, and the three most recent screenshots. This gives the model enough context to understand where it is in a multi-step task, remember what it has already done, and make informed decisions about what to do next — all without requiring an external memory system or state management layer.

How Microsoft Trained Fara-7B

Training a computer-use agent presents a unique challenge: you cannot simply scrape the internet for labeled data the way you would for a language model or image classifier. There are no large public datasets of humans performing web tasks with annotated screenshots and action sequences. Microsoft's solution to this problem is FaraGen, a sophisticated synthetic data generation pipeline that produces high-quality training trajectories at scale.

FaraGen works by combining a task proposal model, a task-solving model, and a verification system. The process begins with a task proposal model — specifically the Magentic-One Orchestrator, which is Microsoft's existing multi-agent framework — that generates realistic web-based tasks. These tasks are practical and grounded in real-world scenarios, such as finding specific products on e-commerce sites, filling out registration forms, or navigating government portals.

For each proposed task, a task-solving agent executes the task step by step, recording every screenshot and action as it goes. This creates a trajectory — a complete record of what the agent saw and did at each step. After the trajectory is generated, a verification step checks whether the task was actually completed successfully. Only trajectories that pass verification are kept for training.

The economics of FaraGen are particularly noteworthy. According to Microsoft's research, each fully verified trajectory costs approximately one dollar to produce. This is dramatically cheaper than having humans manually create training data, and it allows FaraGen to scale to the massive volumes needed for effective training. The system produced 145,000 trajectories comprising 1 million total action steps, covering 70,000 unique websites and domains. This breadth of coverage is essential for training a model that can generalize across the vast diversity of the modern web.

The training itself is supervised fine-tuning, which means the model learns to imitate the actions recorded in the FaraGen trajectories. Notably, Microsoft did not use reinforcement learning — a common technique in other agentic AI systems. Instead, they relied entirely on the quality and diversity of the synthetic data to shape the model's behavior. The fact that this approach produces such strong results is a testament to the effectiveness of the FaraGen pipeline.

The breadth of the training data is especially important. The 70,000 unique domains span e-commerce platforms, social media sites, government services, news outlets, educational portals, and countless other categories. This diversity is what allows Fara-7B to generalize to entirely new websites it has never encountered during training. A model trained only on Amazon and Google would fail on a government tax portal, but Fara-7B's training covers enough of the web's variety that it can reason about unfamiliar interfaces based on visual patterns it has seen before.

Benchmark Results and Performance

Fara-7B establishes a new state of the art for the 7-billion parameter class on every major web agent benchmark. The results are not just good for a small model — they are genuinely competitive with much larger frontier models.

On WebVoyager, which tests general web browsing and task completion across diverse websites, Fara-7B achieves a success rate of 73.5 percent. This compares favorably to UI-TARS-1.5-7B at 66.4 percent, OpenAI's computer-use-preview model at 70.9 percent, and GPT-4o with Set-of-Mark prompting at 65.1 percent. In other words, a 7-billion parameter open-weight model from Microsoft outperforms OpenAI's proprietary computer-use model on general web tasks.

On Online-Mind2Web, a benchmark specifically designed to test generalization to unseen websites, Fara-7B scores 34.1 percent. This is higher than UI-TARS-1.5-7B at 31.3 percent and GPT-4o SoM at 34.6 percent, while trailing OpenAI's computer-use-preview at 42.9 percent. The strong performance here is particularly significant because it demonstrates that Fara-7B's training with FaraGen produces genuine generalization rather than memorization of specific websites.

Fara-7B also performs strongly on DeepShop, a benchmark focused on complex e-commerce tasks that require multiple steps across shopping websites. The model achieves a score of 26.2 percent, which is competitive with much larger models on this challenging evaluation.

Perhaps most interestingly, Microsoft introduced a brand new benchmark called WebTailBench alongside Fara-7B. This benchmark focuses on what Microsoft calls "tail" scenarios — web tasks that are underrepresented in existing benchmarks because they involve less common interactions, unusual website layouts, or edge cases that most benchmarks do not capture. On WebTailBench, Fara-7B achieves 38.4 percent, which is dramatically higher than UI-TARS-1.5-7B at 19.5 percent and OpenAI's computer-use-preview at 25.7 percent. This nearly doubles the performance of competing 7B models on these difficult scenarios, and it suggests that FaraGen's diverse training data gives Fara-7B a significant advantage on the long tail of web tasks that other models struggle with.

An often-overlooked metric that speaks to Fara-7B's efficiency is the average number of steps per task. Fara-7B averages approximately 16 steps to complete a task, while UI-TARS-1.5-7B takes roughly 41 steps for the same tasks. This means Fara-7B is not just more accurate — it is also more than twice as efficient, completing tasks in fewer actions. This efficiency translates directly into lower latency, reduced compute costs, and a better user experience in real-world applications.

Safety and Guardrails

Any model capable of autonomously clicking buttons and typing text on a real computer needs robust safety mechanisms, and Microsoft has built several layers of protection into Fara-7B.

The most visible safety feature is what Microsoft calls Critical Points. Before performing any action that could have irreversible consequences — such as submitting a form, making a purchase, sending an email, or deleting data — Fara-7B pauses and asks the user for explicit consent. This is not a simple yes-or-no prompt. The model presents the specific action it intends to take and the reasoning behind it, giving the user full visibility into what is about to happen. The user can approve the action, request a modification, or cancel the operation entirely.

This consent mechanism is particularly important because it addresses one of the fundamental risks of agentic AI: the possibility of unintended side effects. An AI agent navigating a web form might accidentally submit incomplete information, click the wrong button, or proceed with an action the user did not intend. By requiring human approval at critical junctures, Fara-7B ensures that the user remains in control even while the AI handles the routine work.

Microsoft also strongly recommends running Fara-7B in a sandboxed environment. This means the model should operate in an isolated browser instance or virtual machine that cannot access sensitive personal data, financial information, or critical system resources unless explicitly authorized. The sandboxing approach limits the potential damage from any mistakes the model might make, providing an additional safety net beyond the Critical Points system.

Every action Fara-7B takes is logged and auditable. This creates a complete record of the model's behavior throughout a task, which is essential for debugging, compliance, and trust. If something goes wrong, developers can trace exactly what the model saw, what it decided to do, and why. This audit trail also makes it possible to identify patterns in model failures and improve the system over time.

For harmful or inappropriate tasks, Fara-7B demonstrates strong refusal behavior. In testing, the model refused to engage with approximately 82 percent of tasks it identified as harmful, deceptive, or otherwise problematic. Microsoft trained this refusal capability using a combination of publicly available safety datasets and internally generated refusal tasks through the FaraGen pipeline. The training specifically teaches the model to recognize when a user's request crosses ethical or safety boundaries and to decline participation without being confrontational.

It is worth noting that no safety system is perfect. Microsoft is transparent about the fact that Fara-7B's safety mechanisms, while robust, are not infallible. The combination of Critical Points, sandboxing, action logging, and refusal training creates a defense-in-depth approach, but users and developers should still exercise appropriate caution when deploying the model in sensitive environments.

Fara1.5 — The Next Generation

Microsoft did not stop at Fara-7B. In July 2026, the research team released Fara1.5, a family of three agentic computer-use models built on the Qwen3.5 architecture. The family includes a 4-billion parameter model, a 9-billion parameter model, and a 27-billion parameter model, providing options across different compute budgets and performance requirements.

The progress from Fara-7B to Fara1.5 is substantial. The Fara1.5-9B model achieves 86.6 percent on WebVoyager and 63.4 percent on Online-Mind2Web, massive improvements over the original Fara-7B's 73.5 percent and 34.1 percent respectively. The largest model, Fara1.5-27B, pushes these numbers even further to 89.3 percent on WebVoyager and 72.3 percent on Online-Mind2Web.

To put the Fara1.5-27B results in context, this model outperforms both OpenAI's Operator product and Google's Gemini 2.5 Computer Use on multiple benchmarks. A 27-billion parameter open-weight model beating proprietary products from the largest AI companies in the world is a remarkable achievement, and it validates Microsoft's approach to synthetic data generation and agentic training.

The Fara1.5 release also brings improvements to the training pipeline itself. The FaraGen system has been refined to produce higher-quality trajectories with better coverage of complex multi-step tasks. The verification step has been made more rigorous, filtering out trajectories that are technically successful but inefficient or suboptimal. These improvements compound across the larger model sizes, allowing Fara1.5 to extract more value from its training data than the original Fara-7B could.

The availability of three model sizes in Fara1.5 also reflects Microsoft's understanding that computer-use agents will be deployed across a wide range of environments. The 4B model is small enough for mobile devices and edge computing scenarios, the 9B model fits comfortably on consumer GPUs, and the 27B model provides maximum performance for server-side deployments. This tiered approach ensures that developers can choose the right balance of capability and resource requirements for their specific use case.

Real-World Applications

The practical applications of Fara-7B and its successors are extensive and span multiple industries.

In enterprise automation, companies spend enormous amounts of time on repetitive web-based tasks — filling out compliance forms, extracting data from government portals, processing vendor applications, and managing internal tool configurations. Fara-7B can automate these workflows by navigating the same web interfaces that human employees use, eliminating the need to build custom API integrations for every system.

For accessibility, agentic computer-use models have transformative potential. Users with motor impairments who find it difficult to navigate complex web interfaces could delegate routine tasks — online shopping, bill payments, form submissions — to an AI agent that performs the physical interactions on their behalf. The Critical Points safety mechanism is particularly valuable here, as it keeps the human in the loop for important decisions while the AI handles the mechanical work.

Software testing is another natural fit. QA teams currently spend significant effort writing and maintaining automated tests that are tightly coupled to specific HTML structures. Fara-7B can perform visual regression testing and end-to-end user flow validation by interacting with the application the same way a user would, catching issues that structural test automation might miss.

In customer support and IT helpdesk scenarios, Fara-7B can guide users through complex web-based processes step by step, or perform the actions directly on the user's behalf with their permission. This reduces the burden on support staff and provides faster resolution for common issues like account setup, password recovery, and portal navigation.

The research and data collection applications are equally compelling. Academic researchers, journalists, and market analysts often need to gather information from multiple websites, cross-reference data, and compile reports. A computer-use agent can automate the tedious data collection portion of this work, allowing researchers to focus on analysis and interpretation rather than manual web navigation.

Getting Started with Fara-7B

Developers who want to experiment with Fara-7B can download the model weights from HuggingFace or Microsoft Foundry. The model runs on consumer-grade hardware — a single NVIDIA GPU with at least 16 GB of VRAM is sufficient for the 7B model, while the larger Fara1.5 variants require proportionally more memory.

The model is distributed with the MIT license, which is the most permissive open-source license available. This means there are no restrictions on commercial use, no requirement to share modifications, and no patent implications. Companies can integrate Fara-7B into commercial products without legal concerns, and researchers can modify the architecture or training approach without license compliance overhead.

Microsoft provides reference implementations of the Playwright-based tool interface that Fara-7B uses to interact with browsers. Developers can use these as a starting point and customize the action space for their specific use cases. Some teams may want to add additional browser actions, integrate with desktop applications instead of browsers, or connect the model to mobile device simulators.

The combination of open weights, permissive licensing, and reasonable compute requirements makes Fara-7B one of the most accessible computer-use AI models available today. Unlike proprietary solutions from OpenAI or Google that require API access and ongoing subscription costs, Fara-7B can be deployed entirely on-premises with no external dependencies, which is particularly important for organizations handling sensitive data.

The Bigger Picture

Fara-7B represents a shift in how we think about AI interfaces. For years, the dominant paradigm has been text-in, text-out — you type a prompt, the model generates a response. Computer-use models like Fara-7B break out of this paradigm by operating directly in the visual interface that humans use every day. This is not just a technical novelty. It has profound implications for what AI can actually do in practice.

The text-in, text-out paradigm limits AI to tasks that can be expressed as natural language interactions. But many real-world tasks are fundamentally visual and interactive — navigating a complex web application, filling out a multi-page form with dropdown menus and date pickers, comparing prices across multiple browser tabs, or following a multi-step tutorial. These tasks are difficult or impossible to describe purely in text, but they are natural for a model that can see the screen and click on things.

By bridging the gap between language understanding and visual interaction, models like Fara-7B open up a vast new territory of practical AI applications. The fact that this capability now fits in a 7-billion parameter model that runs on consumer hardware means that the barrier to entry has dropped dramatically. We are moving from an era where computer-use AI required massive cloud infrastructure to one where it can run locally on your laptop, privately and without latency.

Microsoft's investment in Fara-7B and the Fara1.5 family signals that they see agentic computer use as a core capability for the future of AI. The combination of strong benchmarks, open licensing, reasonable compute requirements, and robust safety mechanisms makes this one of the most practical and accessible entries in the rapidly evolving field of AI agents. Whether you are a developer building automation tools, a researcher exploring agentic AI, or an enterprise looking to reduce manual web-based work, Fara-7B is worth serious attention.

Frequently Asked Questions

What is the difference between Fara-7B and other computer-use models like Claude Computer Use?

The fundamental difference is architectural. Claude Computer Use relies on accessibility trees and structured web representations, which means it needs the underlying HTML structure of a page to function. Fara-7B uses only raw screenshots — pixels on a screen — which makes it more robust across different website technologies and less dependent on how a page is built internally. Fara-7B is also significantly smaller at 7 billion parameters, making it runnable on consumer hardware, while frontier models like Claude require API access and substantial cloud compute.

Can Fara-7B run on my laptop?

Yes, with appropriate hardware. The 7B model requires a GPU with at least 16 GB of VRAM, which covers most modern gaming laptops and workstations. The smaller Fara1.5-4B model has even lower requirements and can run on hardware with 8 GB of VRAM. For server deployments, the larger Fara1.5-27B model provides maximum performance but requires enterprise-grade GPU hardware.

Is Fara-7B safe to use with sensitive websites like banking portals?

Microsoft strongly recommends running Fara-7B in a sandboxed environment and taking advantage of the Critical Points system, which pauses before any irreversible action. For highly sensitive applications like banking, the model should be configured to require explicit human approval for every action, not just critical ones. The action logging system provides a complete audit trail for compliance purposes.

How does FaraGen differ from reinforcement learning approaches?

FaraGen produces training data through a generate-and-verify pipeline rather than through trial-and-error learning. The system generates complete task trajectories, verifies whether each was successful, and uses only the successful ones for supervised fine-tuning. This is more sample-efficient and predictable than reinforcement learning, which requires the model to explore many failing attempts before learning what works. Microsoft chose this approach because it produces more consistent results and is easier to control for safety.

What websites does Fara-7B work on?

Fara-7B was trained on trajectories from 70,000 unique domains spanning e-commerce, social media, government services, education, news, and many other categories. The model generalizes well to unseen websites because it understands web interfaces visually rather than memorizing specific HTML structures. However, it works best on standard modern websites and may struggle with highly unusual or custom interfaces.

What is WebTailBench and why does it matter?

WebTailBench is a new benchmark introduced by Microsoft alongside Fara-7B that focuses on underrepresented web tasks — the long tail of interactions that existing benchmarks do not adequately test. These include unusual form layouts, non-standard navigation patterns, and edge-case interactions. WebTailBench matters because real-world web usage involves far more variety than standard benchmarks capture, and a model that performs well on WebTailBench is more likely to succeed on the diverse tasks users actually need completed.

Author's Note: This article is based on publicly available information from Microsoft Research publications, the official Fara-7B model card, and related technical documentation as of August 2026. Model capabilities, benchmarks, and availability may change. Always check https://www.microsoft.com/en-us/research/ for the latest information about Fara-7B and Fara1.5.