Fara 1.5 - Microsoft's browser-based AI agent model series

What is Fara 1.5?

Fara 1.5 is the latest browser-based AI agent (CUA) model series from Microsoft Research AI Frontiers, including three parameter versions: 4B, 9B, and 27B. Based on a fine-tuned Qwen 3.5, the model uses pure pixel-level screenshot input and outputs mouse and keyboard operations through an observation-thinking-action loop. On the Online-Mind2Web benchmark, version 27B achieved a 72% task success rate, surpassing OpenAI Operator and Gemini 2.5 Computer Use. The model is used with the MagenticLite sandbox browser, supporting complex web page tasks such as form filling, cross-site price comparison, and event booking, proactively requesting user confirmation before key operations.

Main functions of Fara 1.5

• Pixel-level browser controlIt directly reads browser screenshots, without relying on the DOM or accessibility tree, and outputs pixel-level mouse and keyboard operations such as click, type, scroll, and drag.
• Webpage-specific shortcutsSupports high-level web page operations such as web_search, visit_url, and go_back, improving task execution efficiency.
• Context management meta-actionsBuilt-in memorize, ask_user, finish and other actions, supporting long-term task memory and collaborative human-computer interaction.
• Security Sandbox ExecutionIt runs through the MagenticLite sandbox browser, forming a security boundary with the user's device, and all operations are traceable.
• Multi-scale deploymentIt offers three versions: 4B, 9B, and 27B, catering to both lightweight deployment of edge devices and high-performance scenarios.

Technical Principles of Fara 1.5

• InfrastructureUsing Qwen3.5 as the base checkpoint, it was supervised fine-tuned (SFT) with approximately 2 million valid samples.
• Observation-Reflection-Action CycleEach input step includes the complete dialogue history and the three most recent browser screenshots. The model first outputs the reasoning process (thought) and then predicts the next atomic action (action).
• Sliding window loss maskCross-entropy loss is calculated only for the thought and action tokens from the most recent 3 rounds; historical screenshots and earlier steps are only used as input context.
• FaraGen 1.5 Synthetic Data EngineTrajectories are generated through multiple rounds of interaction between the teacher agent (GPT-5.4 + Fara1.5 tool capability constraints) and the user simulator. The trajectory is then filtered by a triple validator of correctness, efficiency, and user interaction to supplement data that cannot be collected from real web pages, such as login status and irreversible operations.
• Training data allocationWebpage trajectories 60%, synthetic environment 12.8%, form filling and user interaction 12.5%, fact anchoring 8.8%, visual question answering 4.9%.

How to use Fara 1.5

• Environmental preparationInstall the MagenticLite sandbox browser, an interface designed specifically for Fara 1.5 series agents, providing an isolated execution environment.
• Loading ModelSelect the Fara1.5-4B, 9B, or 27B model based on hardware and performance requirements, and complete the model loading and browser instance initialization.
• Assign task: Describe the web page task using natural language, such as "Book a hotel room in NYC from November 24-28".
• Execute in a loopThe model automatically enters the Observe-Think-Act loop, reads screenshots, infers, and executes single-step actions until the task is completed or the user is actively asked at key nodes.
• Confirmation and InterventionWhen encountering missing information, ambiguous task descriptions, or irreversible operations, the model will pause and request user confirmation to ensure operational safety.

Fara 1.5's core advantages

• Leading performance in its sizeFara 1.5-9B achieved a success rate of 63.4% on Online-Mind2Web, nearly 15 percentage points higher than the similarly sized GUI-Owl-1.5-8B (48.6%), setting a state-of-the-art (SOTA) record for this parameter level. Fara 1.5-4B, with a score of 57.3%, is nearly double that of its predecessor, Fara-7B (34.1%), proving that small models can also handle complex browser tasks.
• Purely vision-driven, with zero external dependenciesIt does not parse the DOM tree, does not rely on accessibility interfaces, and understands the page only through pixel screenshots, making it more adaptable and not limited by changes in web page structure or front-end frameworks.
• Breaking the boundaries of public web pagesBy using the FaraGen 1.5 synthetic data engine to simulate scenarios such as email, calendar, and media platforms that require login or involve irreversible operations, the model's capabilities are expanded from "read-only browsing" to "write-enabled interaction".
• Can be deployed on modest hardwareVersions 4B and 9B can run on ordinary consumer hardware, lowering the barrier to entry for enterprises and developers, while version 27B is designed for scenarios that pursue ultimate performance.

Fara1.5 project address

• Project address: https://www.microsoft.com/en-us/research/articles/fara1-5-computer-use-agent/
• GitHub repositoryhttps://github.com/microsoft/fara

Comparison of Fara 1.5 with similar competing products

Application scenarios of Fara 1.5

• Automated form fillingAutomatically identify fields and fill in information on long forms such as recruitment, expense reimbursement, and appointments; proactively ask the user when a field is missing.
• Cross-site price comparison and researchAutomatically redirects users across multiple e-commerce or information sites, retrieves product parameters and prices, and summarizes and compares the results.
• Schedule and Booking OperationsComplete the booking process for hotels, flights, meeting rooms, etc., which requires multiple clicks, date selection, and payment confirmation.
• Email and Content ManagementWith the support of the synthesis environment, it can perform operations involving account permissions, such as sending emails, saving drafts, and uploading attachments.
• AccessibilityProvides natural language-driven web navigation and interaction agents for visually impaired or operationally limited users.