NorthCUA: The Free Open-Source AI That Uses Your Computer for You

Your cursor is about to stop waiting for you. On November 3, 2026, BEIPO AI published NorthCUA, a family of three open-weight models built for something most AI models cannot do: actually use your computer. They look at screenshots, then click, type, scroll and drag their way through desktop apps and websites: a helper that opens your file manager, a data assistant that fills a spreadsheet cell by cell. NorthCUA is free to download under the permissive Apache-2.0 license, and its official harness ships alongside it as open source. Below: what the three models are, what the benchmarks really say, how to run the whole stack yourself, and the caveats the project's own documentation insists you read.

What NorthCUA actually is

A "computer-use agent" takes your instruction plus screenshots plus interaction history, and outputs tool calls a runtime executes: clicks, keystrokes, drags. The family has three members, all verified live on Hugging Face on November 4, 2026:

All three are fine-tunes on Qwen3.5, released under Apache-2.0, the license that permits commercial use without copyleft conditions. The family page is at huggingface.co/beipo-ai, with full cards for Pro, Mini and Nano. The project's own README stresses the model is only half the system; the other half is the official NorthCUA Harness, which turns tool calls into real actions.

What the numbers say, and who measured them

The project reports its results in a 121-page technical report, published as arXiv 2611.02102. The headline figures for NorthCUA-Pro: 93.30% on ScreenSpot-Pro (GUI grounding accuracy), 90.96% on OSWorld-G, 88.64% on OSWorld-Verified (desktop task success) and 73.89% on WebArena-Verified (web task success). Nano lands 78.90% and 67.16% on the two task benchmarks; Mini sits between. There is one piece of independent corroboration for the grounding score: the November 4, 2026 edition of the Hugging Face Official Benchmarks daily update lists beipo-ai/NorthCUA-Pro as the new #1 on ScreenSpot-Pro at 93.3, with the caveat that most leaderboard entries are self-reported by model authors.

Every one of these figures is company-reported, and the model's own card is unusually upfront about it: it states the research results "have not been tied to an exact checkpoint975 export and are not new measurements of the files in this repository," and tells readers to compare them "with those conditions in view." Two of those conditions matter. OSWorld-G used the best of three complete runs, not an average. And OSWorld-Verified's 88.64% was measured with a 100-step task limit on 361 tasks, excluding eight Google Drive tasks that need separate account setup.

Trained on experience, not just instructions

The idea BEIPO AI announced on November 3 is an "experience flywheel": generate tasks, record the decisions taken, curate the verified trajectories, and train on those. The report's rounded training summary, quoted in the model card: more than 100,000 trajectories and 4 million interaction steps of broad GUI mid-training, about 40,000 grounding examples, roughly 3,000 verified long-horizon trajectories (310,000 GUI steps), and a prepared pool of about 1,500 desktop plus 500 web tasks for joint desktop-web reinforcement learning. The card cautions that these are corpus summaries from the report's Table 3, not per-model consumption counts. Reusing experience in training does not mean the serving model updates its weights while handling your request. The card states that explicitly.

How to run it: the harness is the product

First, use the right repository. The model cards still link the harness at china-qijizhifeng/NorthCUA-Harness, but that address is now an archived migration notice. The maintained code lives at github.com/beipo-ai/NorthCUA (Apache-2.0, Python, created November 2 and updated November 3, 2026). The documented path, from the project's first-run guide:

1. Python 3.12, Git, and the uv package manager.
2. git clone --branch main https://github.com/beipo-ai/NorthCUA.git
3. uv sync --locked --python 3.12 --extra local --extra device, then install Chromium for browser tasks with uv run --no-sync playwright install chromium.
4. northcua --help, then either connect your own model endpoint (model ID plus API key, OpenAI-compatible base URL; it must accept images and structured tool calls) or run the credential-free tutorial first. The tutorial drives real tools against a scripted model endpoint; the docs say plainly it "does not measure autonomous model quality."

The contract between model and harness is strict: every click or drag coordinate is expressed in a normalized 0–1000 space, and the harness converts to pixels at execution time. Mixing in raw pixel coordinates will misplace actions. For the weights themselves: hf download beipo-ai/NorthCUA-Pro --local-dir ./NorthCUA-Pro, then shasum -a 256 -c checksums.sha256. Pro ships 122 Safetensors shards totaling about 793.6 GB, with a 262,144-token configured context length. Nano is the practical starting point for a single-GPU machine.

What it costs

Nothing for the software. There is no published price because there is nothing to buy. Apache-2.0 means you can use the weights commercially; the model's own docs frame the real budget as yours to bear: the GPU (or endpoint) you run it on, the API bills if you point the harness at a hosted model instead of self-hosting, and the disk space (about 794 GB for Pro). That is the complete pricing picture: free weights, your own infrastructure.

The honest caveats

It is a research release, not a product. The model card describes NorthCUA as "intended for research and development of computer-use agents" and requires that you "verify consequential actions and final artifacts, isolate the execution environment, and protect credentials from observed content." Screen data leaves your machine to whatever model endpoint you use; nothing about this stack is local-only unless you serve the weights yourself.

Read the fine print in the first-run guide before installing: public NexAU currently requires an affected Transformers 4.x version, so the docs tell you to read the retained-dependency-risk note first, and to put only trusted model and tokenizer configurations in that environment.

The ecosystem is small and young. The harness repository shows 39 stars; the Pro model card shows 13 likes, both verified live on November 4, 2026. Treat star counts as attention, not reliability.

Where NorthCUA fits

The computer-use lane filled up fast this week. Quick orientation:

NorthCUA is the open-weights desktop option in this lineup: no vendor endpoint, no per-task billing, but also no product support. If you want the agent that lives on your phone instead, start with our Mika review; for the Mac-native alternative, see Sentient OS.

Is NorthCUA really free?

Yes. All three models and the harness are Apache-2.0. There is no published price list because nothing is sold. Your costs are compute, storage and any model-endpoint bills if you do not self-host.

Which model should I start with?

Nano: 9B dense, the cheapest to serve and the documented way to try the stack. Pro's 397B weights need about 794 GB on disk and multi-GPU serving.

How is it different from a chatbot that can browse?

It grounds actions to the screen: every click is a coordinate the model must predict in a normalized 1000-space, and it is trained with long-horizon reinforcement learning across tasks of up to 100 steps, not single-turn answers.

Can I use it without the harness?

The standalone Python example in the model card only demonstrates image-text generation. For computer-use tasks, the official NorthCUA Harness is required.

Should I let it type my passwords?

Not on your daily machine. The project's own guidance: verify consequential actions, isolate the execution environment, and protect credentials from observed content. This is a research release.

Sources

• NorthCUA-Pro model card, Hugging Face — https://huggingface.co/beipo-ai/NorthCUA-Pro (model sizes, license, benchmark table, caveats; verified live November 4, 2026)
• NorthCUA-Nano and NorthCUA-Mini, Hugging Face — https://huggingface.co/beipo-ai/NorthCUA-Nano and https://huggingface.co/beipo-ai/NorthCUA-Mini (9B dense / 35B-3B MoE, Apache-2.0, updated November 3, 2026; verified live)
• NorthCUA Harness, GitHub — https://github.com/beipo-ai/NorthCUA (install docs, first-run guide, technical-report link; read live November 4, 2026)
• Harness migration notice — https://github.com/china-qijizhifeng/NorthCUA-Harness (archived; points to the maintained repository)
• "NorthCUA Technical Report: Turbocharging Computer-Use Agents through Experience at Scale" — https://arxiv.org/pdf/2611.02102v1 (arXiv 2611.02102; training scale and evaluation settings)
• Project homepage — https://beipo.top/blog/northcua/ (official project page)
• KBlip releases — https://kblip.com/releases (November 3, 2026 release note: experience flywheel, Nano/Mini/Pro, weights + code + paper public)
• Hugging Face Official Benchmarks daily update, November 4, 2026 — https://dev.to/quantid/hugging-face-official-benchmarks-daily-update-2026-11-04-20mo (NorthCUA-Pro new #1 on ScreenSpot-Pro at 93.3)