Mage-VL-Cowork

Mage-VL-Cowork

Operates a computer from screenshots and terminal commands with one small vision-language model of about 4B parameters, made by Microsoft for people who build or study computer-use 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: Mage-VL-Cowork is a vision-language model of about 4B parameters from Microsoft. It reads a screenshot and decides whether to click on the screen or run code in a terminal. It was trained from an intermediate Mage-VL checkpoint and handles both kinds of action in one policy, with no outside router. Images and text go in, and text comes out.

Who it suits: It suits researchers and engineers who build or study computer-use agents, and teams that want a small model for on-device work.

Industry | Who uses it | How they use it | Example
AI research | Agent researchers | Study how one small policy chooses between clicking and coding | Compare it with larger computer-use models on desktop tasks
Software | Agent developers | Build desktop assistants that edit files and drive apps | A prototype that fills a spreadsheet through code, then checks the screen
Education | Lab instructors | Teach how computer-use agents are trained | A class reads the model card and technical report together
IT automation | Automation teams | Try a small local model on office workflows | Trial edits of documents, spreadsheets and presentations

Anyone who only needs a chat model or an image captioner can skip it.

Where to use it:

• Python projects that load models with the transformers library
• Windows, macOS or Linux machines, where you download the files
• Agent setups that give a model a screenshot tool and a terminal tool

Use case | Who | What they get back
Desktop task automation | Agent developers | Click and typing actions chosen from each screenshot
Spreadsheet and document edits | Developers | Terminal commands that change files directly instead of many clicks
Small-model research | Researchers | A 4B model to compare against larger computer-use models

How to set it up: You need these first:

• Python and pip
• Disk space for the weights, about 9.5 GB in BF16 or about 2.4 GB at 4-bit as an estimate
• A wrapper program that feeds the model screenshots and runs its actions, since the model itself only produces text

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Load the model with transformers as the model card describes.

pip install -U huggingface_hub
hf download microsoft/Mage-VL-Cowork

How to use it: Each turn you give the model a screenshot and the task. It answers with an action in one of two tools, and your wrapper program carries the action out. The model card holds the exact way to run it.

Tool | What it does
computer_use | Acts on what is visible, such as layouts, canvas editing and live application state
bash | Runs code in a terminal for spreadsheets, documents and presentations, and for checking results

Example: Ask it to "add a totals row to the spreadsheet that is open". From the card's description, it would likely write the cells with a terminal command, check the result on a new screenshot, and click only where the screen alone shows the answer.

For newcomers and experts: It is a small model that can look at a computer screen and also type commands. A first small thing to try is the two download commands above, then reading the loading example in the model card.

Under the hood, training went through five stages that move from separate skills to one unified policy. Before rollouts the teacher model had its tool set trimmed. A synthesis engine called Mage-CUA-Forge then produced 293,806 tasks that cover 6,915 atomic skills. Online reinforcement learning added 10.16 points over the supervised hybrid stage and cut model calls by 13.0%. The reported scores use a 50-step budget.

Task and screenshot go in
Model picks the click or terminal tool
The action runs on the computer
The next screenshot or output comes back

Good to know: The scores below come from the project's own reports on OSWorld-Verified.

Title: OSWorld-Verified success rate at a 50-step budget
Overall: 66.0%
Operating System tasks: 87.5%
Multi-Application Workflows: 62.7%

• Size: Hugging Face lists 4.7B parameters, stored as Safetensors, and the model works in English
• Other versions: 7 quantized and 2 fine-tuned versions made by other people are listed
• License: Apache 2.0
• Cost to run: none named, you supply the hardware
• Data it reaches: the screenshots you give it and whatever machine or terminal you connect it to

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 4.7B
• Weights: about 9.5 GB in BF16; about 2.4 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 4,226
• Likes: 222
• License: apache-2.0
• Languages: en
• Based on: microsoft/Mage-VL
• Quantized versions: 8
• Fine-tunes: 2
• Tags: transformers, safetensors, llava_onevision2, multimodal, vision-language-model, computer-use, hybrid-agent, gui-agent, coding-agent, terminal-agent, reinforcement-learning, osworld

Links

• Model: https://huggingface.co/microsoft/Mage-VL-Cowork
• Files: https://huggingface.co/microsoft/Mage-VL-Cowork/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download microsoft/Mage-VL-Cowork

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