NorthCUA logo, the open-weight computer-use model family and agent harness

NorthCUA: Power-User Guide to the Open-Weight Computer-Use Models and Harness

by

Last updated: November 4, 2026

NorthCUA is a new open-weight family of computer-use models, Nano (9B), Mini (35B MoE) and Pro (397B MoE), released on Hugging Face from October 30 to November 1, 2026 under Apache-2.0. It ships with the NorthCUA Harness, which lets a vision model control a desktop or browser. Pro reports 88.64 on OSWorld-Verified.

Contents

A computer-use model looks at screenshots and returns actions such as clicks, typing and scrolling. NorthCUA is published by the beipo-ai organization on Hugging Face together with a 121-page technical report and an open-source Harness repository. This guide covers what was released, what the published numbers mean and how to run a first task, using only the official model cards, report and documentation.

What NorthCUA includes

Component Details
Models NorthCUA-Nano, NorthCUA-Mini and NorthCUA-Pro
Base models Qwen3.5-9B, Qwen3.5-35B-A3B and Qwen3.5-397B-A17B (an independent release, not an official Qwen model)
Harness Open-source Python runtime with desktop and web tools, context compaction, working notes, an independent evaluator, batch runs, resume and retries
License Apache-2.0 for the model weights and documentation
Report NorthCUA Technical Report, arXiv 2611.02102
Training data (report) Over 100,000 trajectories and 4 million steps in mid-training, about 40,000 grounding examples, then about 3,000 verified trajectories (310,000 GUI steps) for supervised fine-tuning
Reinforcement learning Joint desktop and web GRPO on about 1,500 desktop and 500 web tasks

Model sizes and hardware

According to the NorthCUA-Pro model card and the NorthCUA-Mini model card, all three models use full Safetensors weights, mostly in BF16, with a configured context of 262,144 tokens. Active parameters reduce compute per token, but you still need memory for every weight.

Model Parameters Transformers class Shards Files on disk
NorthCUA-Nano 9B dense Qwen3_5ForConditionalGeneration 4 About 18.84 GB
NorthCUA-Mini 35B total, 3B active (MoE) Qwen3_5MoeForConditionalGeneration 16 About 70.23 GB
NorthCUA-Pro 397B total, 17B active (MoE) Qwen3_5MoeForConditionalGeneration 122 About 793.6 GB

The Pro card notes that the export records Transformers 5.20.2, that the released checkpoint is release r13 checkpoint 975, and that it does not include multi-token prediction weights.

Reported benchmarks

The technical report gives these scores. Grounding results come from the fine-tuned models, and agent results come from the reinforcement learning stage. OSWorld-Verified uses 361 tasks with a 100-step limit, and WebArena-Verified uses 812 tasks.

Benchmark Nano Mini Pro
ScreenSpot-Pro 72.87 92.41 93.30
OSWorld-G 82.98 87.94 90.96
OSWorld-Verified 78.90 85.30 88.64
WebArena-Verified 67.16 70.43 73.89

Reinforcement learning lifted Pro on OSWorld-Verified from 85.87 to 88.64. On the longer OSWorld 2.0 set (108 tasks, 500-step limit), Pro averaged 66.80% partial credit. The model card states these are self-reported numbers for the report’s checkpoints and were not measured on the exported checkpoint, so treat them as a guide rather than a guarantee.

How to download the weights

  1. Install or update the Hugging Face client: pip install -U huggingface_hub.
  2. Download a model, for example hf download beipo-ai/NorthCUA-Pro --local-dir ./NorthCUA-Pro. Replace Pro with Mini or Nano for the smaller models.
  3. Pin an exact version for reproducible work by adding --revision with the commit hash listed in the documentation.
  4. Verify the files with shasum -a 256 -c checksums.sha256 inside the download folder.
  5. Serve the model with an inference engine that supports the Qwen3.5 architecture and exposes an image-capable, tool-calling API.

The models documentation notes that the beipo-ai links currently resolve to the china-beipo resources on Hugging Face, and that the repository does not claim a tested server-launch recipe for Pro. Our Hugging Face hf CLI guide covers download options, and the vLLM 0.32 guide covers a common serving engine.

How to set up the NorthCUA Harness

The Harness README requires Python 3.12, Git and uv 0.13.6 or newer. A no-key tutorial lets you test the runtime before connecting any model.

  1. Clone the repository: git clone https://github.com/beipo-ai/NorthCUA.git, then change into the folder.
  2. Install dependencies: uv sync --locked --python 3.12 --extra local --extra device.
  3. Install the browser used for web tasks: uv run --no-sync playwright install chromium.
  4. Confirm the install with uv run --no-sync northcua --help.
  5. Run the no-key tutorial: northcua demo serve --directory logs/tutorial, then validate it with northcua config validate and run northcua local web run with --max-steps 6 --evaluate.
  6. Connect your own model: northcua config init --directory logs/my-config --model "$NORTHCUA_MODEL" --base-url "$NORTHCUA_BASE_URL", then validate logs/my-config/app.example.yaml.
  7. Try a bounded desktop task, such as northcua local desktop run with --query 'Open a calculator and compute 12 * 7.' --max-steps 15.

The Harness supports four API interfaces, selected with --api-type: openai_chat_completion, openai_responses, anthropic_chat_completion and gemini_rest. The CLI does not load .env automatically, so export the variables in your shell or pass --model and --base-url directly.

Coordinate settings and other tips

  • Match the coordinate format: NorthCUA models output positions on a 0 to 1000 scale on both axes, where (500, 500) is the screen center. The Harness default is image pixels, so add desktop: coordinate_policy: input_space: normalized_1000 to your app YAML when you use a NorthCUA model.
  • Start small: the documentation recommends one bounded task and a review of the actual tool calls before you run batches.
  • Record the revision: log the checkpoint commit with your results, since published scores belong to specific checkpoints.
  • Use the operator skill: the repository includes skills/northcua/SKILL.md to help a coding agent drive the Harness.
  • Keep secrets out of commits: --api-key-env takes the name of a variable, never the key itself, and local configs and raw traces should not be committed.

How it compares

Hosted options such as the Gemini API Computer Use tool run on the vendor’s servers and bill per token. NorthCUA gives you the weights, so you control hosting, data and cost, but you also carry the hardware and serving work. Tencent Hy4 is another recent model with computer-use features, covered in our Tencent Hy4 guide. For a desktop agent app rather than a model, see our Claude Duet guide.

Limitations

  • The models are intended for research and development. Validate actions before running them on systems that matter.
  • Screenshots can carry prompt injection, so a web page or document can try to steer the agent.
  • The model card notes that one control example it shows was seen during training.
  • Benchmark scores are self-reported and were not measured on the exported checkpoint.
  • The NorthCUA-Ultrafast results in the report do not apply to the released weights.
  • The Harness is built on NexAU, which requires Transformers 4.x, so keep it in a separate environment from the model server.

Frequently asked questions

What is NorthCUA?

NorthCUA is a family of open-weight computer-use agent models, NorthCUA-Nano, NorthCUA-Mini and NorthCUA-Pro, plus the NorthCUA Harness, an open-source runtime that lets a vision model operate a desktop or web browser through tool calls. The weights and documentation are released under Apache-2.0.

How big is NorthCUA-Pro?

NorthCUA-Pro is a mixture-of-experts model with 397 billion total and 17 billion active parameters, built on Qwen3.5-397B-A17B. Its BF16 weights are split into 122 Safetensors shards and take about 793.6 GB on disk, with a configured context length of 262,144 tokens.

What benchmark scores does NorthCUA report?

The technical report lists NorthCUA-Pro at 88.64 on OSWorld-Verified and 73.89 on WebArena-Verified after reinforcement learning, and 93.30 on ScreenSpot-Pro. These are self-reported results for the report’s checkpoints, and the model card says they were not measured on the exported checkpoint.

Can I run NorthCUA locally?

Yes, with suitable hardware. The Nano model needs about 18.84 GB of disk and the Mini about 70.23 GB, while Pro needs about 793.6 GB. You serve the model separately with an inference engine that supports the architecture, then point the NorthCUA Harness at that endpoint.

Does the NorthCUA Harness work with other models?

Yes. The Harness does not require NorthCUA weights. It can connect to any model that accepts image inputs and structured tool calls through OpenAI-compatible Chat Completions, OpenAI Responses, Anthropic Messages or Gemini REST interfaces.

What license does NorthCUA use?

All three model cards and the checkpoint license files specify Apache-2.0, and the Harness repository is open source. The project describes the models as intended for research and development, and asks users to validate actions before running them on real systems.

Sources

Image: NorthCUA logo from the NorthCUA-Pro Hugging Face repository, Apache-2.0 (resized). FreeAnswers is not affiliated with or endorsed by the NorthCUA project.



Related answers

Leave a Comment