microsoft/Fara1.5-27B · Hugging Face

Source: https://huggingface.co/microsoft/Fara1.5-27B

Fara1.5-27B is a multimodalcomputer use agent (CUA)for web browsers, fromMicrosoft Research AI Frontiers. It observes the browser through screenshots and acts on the user’s behalf by emitting structured tool calls — click, type, scroll, visit URL, web search, and so on — to complete tasks end-to-end.

The model is vision-only at perception time: it sees the browser through screenshots, not the DOM or accessibility tree. Internal reasoning and trajectory history are tracked as text. Given the latest screenshot and prior actions, it predicts the next action with grounded arguments (e.g., pixel coordinates for a click).

Fara1.5-27B is supervised fine-tuned from Qwen3.5-27B on data generated byFaraGen1.5, our multi-agent pipeline that synthesizes web tasks, executes trajectories to solve them, and verifies the results before training.

It’s co-designed withMagenticLite, and that’s the recommended deployment for both research and production.

https://huggingface.co/microsoft/Fara1.5-27B#highlightsHighlights

• **End-to-end web task completion.**Fills forms, books reservations, applies for jobs, plans trips, runs shopping carts. Not just clicking around — sequencing actions toward a goal.
• **Vision-only perception.**Operates on screenshots alone, no DOM access required. Matches the input modality available to a human user.
• **Coordinate-grounded actions.**Predicts pixel-level click and drag targets directly. No separate grounding model needed.
• **Critical-points safety design.**Trained to stop and ask before personal info entry, payments, submissions, sign-ins, sending messages, or other irreversible actions — even if it could technically continue.
• **262K context.**Long enough for multi-screenshot trajectories with full action history.
• **On-device viable.**27B parameters means it runs on a single A100/H100/B200 with room for the screenshot history.

https://huggingface.co/microsoft/Fara1.5-27B#model-detailsModel Details

DeveloperMicrosoft Research AI FrontiersArchitectureMultimodal decoder-only LM (image + text → text)Parameters27BContext length262,144 tokensInputsUser goal (text), current screenshot(s), prior agent thoughts and actionsOutputsChain-of-thought block followed by a tool-call block (XML-tagged)Training periodJanuary 2026 – April 2026Training compute64 × NVIDIA B200, 6 daysRelease date21 May 2026LicenseMITBase modelQwen3.5-27B

https://huggingface.co/microsoft/Fara1.5-27B#recommended-deployment-magenticliteRecommended Deployment: MagenticLite

The safest way to run Fara1.5-27B is insideMagenticLite, which provides:

• Sandboxing— the browser runs in a Docker container with no access to host files or environment variables
• Allow-lists— restrict navigation to a user-specified set of domains
• Watch-mode— real-time monitoring of every action with full trace logs
• Pause— immediate halt of agent activity at any point

If you integrate Fara1.5-27B directly, you’re responsible for these controls. Don’t run this model with unrestricted browser access on a machine that has anything sensitive on it.

https://huggingface.co/microsoft/Fara1.5-27B#quickstartQuickstart

https://huggingface.co/microsoft/Fara1.5-27B#requirementsRequirements

• torch \>= 2\.11\.0
• transformers \>= 5\.2\.0
• vllm \>= 0\.19\.1
• A set of GPUs with enough memory for a 27B model in bf16 (A6000, A100, H100, and B200 have been tested). We recommend sharding the model over at least 2 GPUs.

https://huggingface.co/microsoft/Fara1.5-27B#serve-with-vllmServe with vLLM

vllm serve microsoft/Fara1.5-27B \
  --dtype bfloat16 \
  --max-model-len 262144 \
  --limit-mm-per-prompt image=10

https://huggingface.co/microsoft/Fara1.5-27B#system-promptSystem prompt

Fara1.5-27B is trained against a specific system prompt. Use it verbatim for best results:

You are Fara, a computer use agent (CUA) specialized for web browsers. You are developed by Microsoft AI Frontiers. You assist users with completing and automating tasks that require the use of a web browser.

The model was trained in the timeframe of January - April 2026. You can effectively perform tasks even beyond this range by accessing the web browser and using the latest information on the live web. But your knowledge cutoff is limited to early 2026, so you may not be aware of events or developments that occurred after that time, without explicitly browsing and searching for latest information on the web.

This edition of the model was trained using SFT on top of Qwen3.5-27B, using a synthetic data mixture generated and developed by Microsoft AI Frontiers.

A critical point is a situation where we must pause and request information or confirmation from the user before proceeding. There are three types:

Case 1: Missing User Information — The task requires personal information that the user has not provided (e.g., email, phone number, address, payment details). Never fabricate or assume personal information. Fill in only what the user has explicitly provided, then pause and ask for any missing required fields.

Case 2: Underspecified Task — The task description is ambiguous or missing details needed to make a decision at the current step. Pause and ask for clarification.

Case 3: Irreversible Action — We are about to perform an action that cannot be undone (e.g., submitting a form, completing a purchase, sending a message, deleting data). If the user explicitly authorized the action, proceed. Otherwise, stop and ask for confirmation.

Only stop at a critical point if (1) required information is missing, (2) the task is ambiguous, OR (3) an irreversible action lacks explicit user authorization.

The full system prompt, including the completecomputer\_usetool schema, ships with the model in MagenticLite.

https://huggingface.co/microsoft/Fara1.5-27B#tool-schemaTool schema

Fara1.5-27B emits actions as<tool\_call\>\.\.\.</tool\_call\>XML blocks containing a JSON object that calls thecomputer\_usefunction. Supported actions:

ActionPurposeleft\_click,right\_click,double\_click,triple\_clickMouse clicks at\(x, y\)``mouse\_move,left\_click\_dragCursor positioning and dragtype,keyKeyboard inputscroll,hscrollPage scrollingvisit\_url,history\_backBrowser navigationweb\_searchSearch querypause\_and\_memorize\_factPersist a fact across the trajectoryask\_user\_questionSurface a clarifying question to the userwaitSleep for N secondsterminateEnd the task with a final answer The screen resolution Fara is most commonly trained with is1440×900. Match this in your sandbox for the most reliable grounding.

https://huggingface.co/microsoft/Fara1.5-27B#minimal-agent-loopMinimal agent loop

# Pseudocode for a single-step interaction. Use vLLM's OpenAI-compatible
# endpoint and pass the screenshot as an image content part.

screenshot_b64 = capture_browser_screenshot()  # base64 PNG

messages = [
    {"role": "system", "content": FARA_SYSTEM_PROMPT},
    {"role": "user", "content": [
        {"type": "text", "text": "Book a table for 2 at a sushi place in Sunnyvale for Friday 7pm."},
        {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{screenshot_b64}"}},
    ]},
]

response = client.chat.completions.create(
    model="microsoft/Fara1.5-27B",
    messages=messages,
    temperature=0.0,
    max_tokens=2048,
)

# Parse the assistant's reply: it will contain chain-of-thought text
# followed by <tool_call>{"name": "computer_use", "arguments": {...}}</tool_call>.
# Execute the action in your sandboxed browser, capture the new screenshot,
# append the assistant turn and a tool/observation turn, and loop until
# the model emits action="terminate". We only keep the most recent 3 screenshots
# in the chat history.

A reference implementation of the full agent loop, including screenshot capture and Docker sandboxing, is in MagenticLite.

https://huggingface.co/microsoft/Fara1.5-27B#critical-points-safety-by-designCritical Points: Safety by Design

Fara1.5-27B is trained topause and ask the userat three types of critical points:

1. **Missing user information.**If the task needs personal data (name, email, phone, address, payment) the user hasn’t provided, fill in what’s available and stop to ask for the rest. Never fabricate.
2. **Underspecified tasks.**If the goal is ambiguous at the current decision point (e.g., “book me a flight” with no destination), stop and clarify.
3. **Irreversible actions.**Submitting forms, completing purchases, sending messages, deleting data — stop unless the user explicitly authorized the action upfront.

Concrete actions Fara is trained to stop on without explicit authorization:

• Entering personal information (name, email, phone)
• Entering credit card or shipping/billing details
• Completing purchases or bookings
• Making phone calls
• Sending emails
• Submitting job applications
• Signing into accounts

If the user grants authorization, the model proceeds. The principle is that costly, irreversible actions require a human in the loop.

https://huggingface.co/microsoft/Fara1.5-27B#evaluationEvaluation

We evaluate Fara1.5-27B on end-to-end web agent benchmarks. For WebTailBench, we report outcome success.

https://huggingface.co/microsoft/Fara1.5-27B#trainingTraining

https://huggingface.co/microsoft/Fara1.5-27B#approachApproach

Supervised Fine-Tuning on top of Qwen3.5-27B. Training mix is generated by FaraGen1.5, our multi-agent data pipeline, supplemented with curated public datasets for grounding and UI understanding.

https://huggingface.co/microsoft/Fara1.5-27B#data-sourcesData sources

• Synthetic trajectories— Tasks seeded from URLs sampled from a large web index and from open-source seed datasets (e.g., Mind2Web). A multi-agent system attempts each task while recording screenshots, thoughts, and actions. A verifier agent then keeps only successful trajectories.
• Grounding— Curated datasets for predicting actions and pixel coordinates from screenshots, with images, text, and bounding boxes.
• UI understanding— Visual question answering, captioning, and OCR over web page screenshots collected by the pipeline.
• Safety and instruction following— Refusal data covering harmful or unsafe tasks the model should decline.

https://huggingface.co/microsoft/Fara1.5-27B#scaleScale

• Approximately 1 billion text tokens
• Less than 1 billion training images
• Latest data acquisition: March 20, 2026
• Training start: January 2026

The corpus is static. Future updates will ship as separately versioned models with their own model cards.

https://huggingface.co/microsoft/Fara1.5-27B#intended-use-and-limitationsIntended Use and Limitations

https://huggingface.co/microsoft/Fara1.5-27B#primary-use-casesPrimary use cases

Automating repetitive web tasks: filling forms, shopping, booking travel, restaurant reservations, information seeking, account workflows. Fara1.5-27B can also serve as a grounding model for other agents that need pixel-accurate action prediction.

https://huggingface.co/microsoft/Fara1.5-27B#out-of-scopeOut of scope

• Languages other than English (training data is English-only)
• High-stakes domains (legal, health, financial advice) where inaccurate actions could cause harm
• Allocation decisions affecting legal status, housing, employment, or credit
• Unsandboxed deployments with access to sensitive accounts or files
• Commercial or real-world production use without additional testing and safeguards

https://huggingface.co/microsoft/Fara1.5-27B#known-limitationsKnown limitations

• Vision-only perceptionmeans the model can be misled by deceptive or low-quality page rendering, prompt injections embedded in page content, or visual ambiguity in UI elements
• Multi-step trajectories accumulate error— a misclick early in a sequence can compound
• Run-to-run varianceon multi-turn tasks is non-trivial; benchmark numbers are averaged over multiple runs
• The model can hallucinate page state or misattribute information from earlier screenshots

https://huggingface.co/microsoft/Fara1.5-27B#responsible-ai-considerationsResponsible AI Considerations

Computer use is a powerful capability. An agent that can click, type, and submit in a real browser can also do those things wrong. We strongly recommend:

• Human-in-the-loop monitoringof Fara’s actions on the live web with a fast way to halt execution
• Sandboxed execution— run Fara in an isolated container with no access to host files, environment variables, or sensitive credentials
• Allow-listed or block-listed browsing— restrict the agent’s reachable surface to limit exposure to malicious pages
• No credential or PII sharingwith the agent unless absolutely required and the user has authorized it
• Output verification— Fara can hallucinate, misattribute sources, or be misled by deceptive content; verify before acting on its outputs

https://huggingface.co/microsoft/Fara1.5-27B#safety-evaluationSafety evaluation

Fara1.5-27B was evaluated via automated red-teaming on Azure. Coverage included groundedness, jailbreak resistance, harmful content (hate, violence, self-harm), and copyright violations. Safety post-training includes both refusal data for malicious tasks and the critical-points framework described above.

https://huggingface.co/microsoft/Fara1.5-27B#known-model-risksKnown model risks

Like all language models, Fara1.5-27B can produce unfair, unreliable, or offensive outputs. Specific concerns:

• Quality of servicevaries across English dialects and is worse for non-English content
• Representation harmsmay persist despite safety post-training due to base-model biases
• Information reliability— generated content may be inaccurate or outdated
• Prompt injection— Fara reads page content, and adversarial content on a page may attempt to redirect its behavior

Developers deploying Fara1.5-27B should apply responsible AI best practices and ensure compliance with applicable laws and regulations. Using safety services likeAzure AI Content Safetyis recommended.

https://huggingface.co/microsoft/Fara1.5-27B#licenseLicense

Released under theMIT License.

https://huggingface.co/microsoft/Fara1.5-27B#contactContact

For information requests under the EU AI Act and related inquiries:[email protected]

Authorized representative: Microsoft Ireland Operations Limited, 70 Sir John Rogerson’s Quay, Dublin 2, D02 R296, Ireland.