NorthCUA-Pro

NorthCUA-Pro

Reads screenshots and requests, then writes reasoning and tool calls for an agent that drives a desktop or browser. For teams with a large GPU setup who build computer-use agents.

Made by beipo-ai 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: NorthCUA-Pro is the largest model in the NorthCUA family and is built on Qwen3.5. It is a mixture-of-experts model with 397B total parameters and 17B active. You feed it a request, the current screenshots and what happened so far. It answers with its reasoning, a reply and the tool calls that drive a desktop or a browser.

Who it suits: It suits teams that build agents that operate software through the screen, and researchers who study such agents.

Industry | Who uses it | How they use it | Example
Software testing | QA engineers | Automate screen tasks in desktop and web apps | An agent that works through a web form
Operations | Automation teams | Carry out multi-step tasks across apps | A task that moves data from one app to another
Research | AI researchers | Study computer-use agents on shared benchmarks | A run on OSWorld-Verified

People with modest hardware can skip it and look at the smaller Nano or Mini models in the same family.

Where to use it:

• The official NorthCUA runtime package, which the card says to use with this model
• A desktop or browser that the runtime controls
• Hugging Face Transformers, the library the card lists
• Windows, macOS or Linux for downloading the files with Python

Use case | Who | What they get back
Desktop tasks | Automation teams | Reasoning, a reply and tool calls for each step
Web workflows | QA engineers | Actions in a browser that follow a screenshot
Screen grounding | Researchers | Click targets as coordinates in a 0 to 1000 space

How to set it up: You need these first:

• Python and the huggingface_hub package
• The official NorthCUA runtime, from the card's code repository
• A large GPU setup, since the weights are about 794 GB in BF16 or about 198 GB at 4-bit by estimate

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Set up the runtime as the card describes and point it at the model.

pip install -U huggingface_hub
hf download beipo-ai/NorthCUA-Pro

How to use it: Give the runtime a request. It sends the model your request, screenshots and history. The model answers with reasoning, a reply and tool calls. All action coordinates use a 0 to 1000 scale, and the runtime turns them into pixels when it acts.

Example: You ask the agent to complete a task in a browser. The model looks at the current screenshot and replies with its next action, such as a click, as coordinates on the 0 to 1000 scale. The runtime carries out the action and sends back a fresh screenshot.

For newcomers and experts: A computer-use agent looks at the screen and clicks and types the way a person would. A first small thing to try is the smaller Nano model, since this one needs a huge setup.

For experts, the 17B active figure counts what is computed for each token. Memory still has to hold all 397B parameters, so size the hardware on the total. The card says grounding results come from SFT checkpoints and task results from later joint desktop and web RL checkpoints. It adds that these figures come from the report and are not new measurements of these files.

You give a request and screenshots
The model plans and writes tool calls
The runtime converts coordinates to pixels
The runtime runs the action
A new screenshot comes back

Good to know:

Title: OSWorld-Verified success rate, as reported
NorthCUA-Nano: 78.90%
NorthCUA-Mini: 85.30%
NorthCUA-Pro: 88.64%

• Size: 397B total and 17B active parameters in Safetensors format, with 1 quantized version made by someone else
• Family: Nano at 9B dense, Mini at 35B with 3B active, and Pro
• License: apache-2.0
• Cost to run: none named, no API key or account, but the hardware is large
• Data it reaches: it downloads model files from Hugging Face and sees the screenshots and history the runtime gives it
• Not for: use without the runtime, which the card says it is designed to work with, or raw pixel coordinates

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 397B
• Weights: about 794 GB in BF16; about 198 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,066
• Likes: 33
• License: apache-2.0
• Quantized versions: 1
• Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, northcua, computer-use, multimodal, agents, conversational

Links

• Model: https://huggingface.co/beipo-ai/NorthCUA-Pro
• Files: https://huggingface.co/beipo-ai/NorthCUA-Pro/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download beipo-ai/NorthCUA-Pro

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