Fara1.5-9B

Fara1.5-9B

A computer use agent model from Microsoft Research that reads browser screenshots and picks the next click, typing or scroll to finish web tasks. It suits developers and researchers building browser agents.

Made by microsoft on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Fara1.5-9B is a 9 billion parameter agent model from Microsoft Research AI Frontiers that drives a web browser. It sees the page only through screenshots, then predicts the next action, such as a click, typing, scrolling, opening a page or a web search, with pixel coordinates where needed. It is fine-tuned from Qwen3.5-9B on tasks that a multi-agent pipeline generated, ran and checked.

Who it suits: It suits developers and researchers who build browser agents and have a GPU to serve a 9B model. The card says it should run only inside its own runner, which has sandboxing and pause controls.

Title: Who might use it
Industry | Who uses it | How they use it | Example
Software | Agent developers | Drive a browser from screenshots alone | Fill in a web form step by step toward a stated goal
Research | Agent researchers | Study computer use agents in a controlled runner | Watch every action in the runner trace logs
Operations | Teams with repetitive web chores | Automate routine browser tasks with approvals | Prepare a booking and stop before the payment step

If you cannot sandbox a browser, you can skip it. The card warns against unrestricted browser access on a machine with anything sensitive on it.

Where to use it: The card says to use it only inside the runner from the Fara command line project or inside MagenticLite, which it recommends.

• MagenticLite, the recommended setup, with a Docker sandbox, domain allow-lists, watch mode and a pause control
• The Fara command line runner
• A vLLM server on a GPU such as an A6000, A100, H100 or B200

Title: Jobs it does
Use case | Who | What they get back
Finish a web task | Agent developers | A sequence of actions toward the goal, such as filling a form or booking
Pause for approval | Operators | A stop and a question before payments, sign-ins or other irreversible steps
Watch an agent run | Researchers | Trace logs of every action

How to set it up: You need recent versions of torch, transformers and vLLM, plus a GPU that can hold a 9B model at bf16 precision. The card trains the model against one specific system prompt that you should use as written.

• torch 2.11.0 or newer
• transformers 5.2.0 or newer
• vLLM 0.19.1 or newer
• A GPU with room for the bf16 weights, about 18.8 GB

1. Install the Hugging Face download tool and fetch the model files.
2. Install the required versions of torch, transformers and vLLM.
3. Serve the model with vLLM.
4. Use the card's system prompt exactly, inside MagenticLite or the Fara runner.

pip install -U huggingface_hub
hf download microsoft/Fara1.5-9B

vllm serve microsoft/Fara1.5-9B \
  --dtype bfloat16 \
  --max-model-len 262144 \
  --limit-mm-per-prompt image=10

How to use it: You give the agent a goal in text. At each step it receives the latest screenshot and its earlier actions, writes its reasoning, then emits a tagged tool call. The runner carries out the action and sends back a new screenshot.

Title: Actions it can take
Tool | What it does
click | Clicks at pixel coordinates it predicts
type | Enters text into a field
scroll | Moves the page up or down
visit URL | Opens a page by its address
web search | Runs a search

Example: You ask the agent to prepare a reservation on a booking site. It looks at each screenshot, clicks and types its way through the form, and then stops to ask you before entering personal details or paying.

For newcomers and experts: If you are new, think of it as a assistant that looks at your browser the way you do and then clicks and types for you. Start only inside the MagenticLite sandbox, with a harmless task.

If you know the field, it perceives only screenshots, with no DOM or accessibility tree, and predicts grounded actions directly, so no separate grounding model is needed. Its context is 262,144 tokens, enough for many screenshots with their history. Training used data from FaraGen1.5, which synthesizes tasks, runs trajectories and verifies results.

A goal and the latest screenshot go in
The model writes its reasoning
It emits a tagged tool call
The runner acts and returns a new screenshot

Good to know: The card trains the model to stop and ask before personal information, payments, submissions, sign-ins, sent messages and other irreversible steps. If you wire it up yourself, you own those controls.

Title: Approximate weight size
BF16: 18.8 GB
4-bit estimate: 4.7 GB

• License: MIT
• Cost to run: none named, though you need a GPU that holds a 9B model in bf16
• Data it reaches: web pages in a browser, seen as screenshots, and the model files from Hugging Face
• Not for: unrestricted browser access on a machine with sensitive data, or use outside the Fara runner or MagenticLite

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 9.4B
• Weights: about 18.8 GB in BF16; about 4.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 3,008
• Likes: 64
• License: mit
• Languages: en
• Based on: Qwen/Qwen3.5-9B
• Quantized versions: 11
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, computer-use, web-agent, multimodal, vision-language, agent, browser-automation, magentic, fara

Links

• Model: https://huggingface.co/microsoft/Fara1.5-9B
• Files: https://huggingface.co/microsoft/Fara1.5-9B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download microsoft/Fara1.5-9B

Then follow the usage in its model card. Nothing in this kit runs by itself.