NorthCUA-Nano

NorthCUA-Nano

Operates desktop and browser apps from screenshots with a 9B dense computer-use model built on Qwen3.5. Suits developers who build computer-use agents with the official NorthCUA runtime.

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-Nano is a 9B dense computer-use model built on Qwen3.5. Given a request, screenshots and the record of past actions, it writes out its reasoning, a reply and the tool calls that a desktop or browser runtime executes. It is the compact member of the NorthCUA family, next to the 35B Mini and the 397B Pro.

Who it suits: Developers and researchers who build agents that click, type and navigate real software, and who can run the official NorthCUA runtime alongside the model.

Industry | Who uses it | How they use it | Example
Software | Agent developers | Drive desktop apps from screenshots | Open a settings dialog and change a value
Web operations | Automation teams | Carry out browser workflows | Fill in a multi-step web form
Research | AI researchers | Study computer-use agents and compare model sizes | Compare Nano with Mini on the same tasks

If you want a plain chat model, you can skip it, since it is meant to run inside the official runtime.

Where to use it:

• The official NorthCUA runtime, which the card says is required for computer use
• Desktop application and browser runtimes
• Windows, macOS or Linux, for downloading the files

Use case | Who | What they get back
Desktop tasks | Agent developers | Actions and tool calls for a desktop runtime
Web tasks | Automation teams | Steps for a browser workflow
GUI grounding | Researchers | Coordinates for on-screen elements

How to set it up:

• Python with pip
• Hardware for about 18.8 GB in BF16, or about 4.7 GB at 4-bit by Hugging Face's estimate
• The official NorthCUA runtime, which carries the prompts, tool calls and action execution

1. Install the Hugging Face hub package.
2. Download the model files.
3. Set up the official runtime and connect it to the model.

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

How to use it: Give the runtime a task and let it pass screenshots and history to the model. The model returns actions, and the runtime performs them. All GUI coordinates use a normalized space from 0 to 1000, not raw pixels, and the runtime converts them at execution time.

Tool | What it does
NorthCUA runtime | Handles prompts, tool calls, observations and action execution
normalized_1000 | The coordinate convention, 0 to 1000 across the screen

Example: You ask the agent to change a setting in a desktop app. It looks at a screenshot, decides which element to click, returns the action with coordinates in the 0 to 1000 space, and the runtime performs the click.

For newcomers and experts: If you are new, treat this as a research-grade agent and start in a safe test environment, since it operates real software. Read the runtime setup before anything else.

The card reports results from its technical report, not new measurements of these files. Grounding scores come from supervised checkpoints and task completion from later reinforcement learning checkpoints. Nano is the dense model, so all 9B parameters are active for each token.

A task and a screenshot arrive
The model reasons and picks an action
Coordinates use the 0 to 1000 space
The runtime converts them to pixels
The action runs on the desktop or browser

Good to know:

• The card says the model and the runtime are designed to be used together
• The reported scores are research results, not new measurements of this repo
• License: Apache 2.0
• Cost to run: none named, you supply the hardware
• Data it reaches: screenshots and interaction history from the desktop or browser it controls, plus the files downloaded from Hugging Face
• Not for: raw pixel coordinates, since all actions use the normalized convention

Title: Reported scores for Nano
ScreenSpot-Pro: 72.87 percent
OSWorld-G: 82.98 percent
OSWorld-Verified: 78.90 percent
WebArena-Verified: 67.16 percent

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: 485
• Likes: 14
• License: apache-2.0
• Quantized versions: 2
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, northcua, computer-use, multimodal, conversational

Links

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

Get the model

Download its files with Hugging Face's command line:

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

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