NorthCUA open-source AI agent that uses your computer - Nano, Mini and Pro models

The Open-Source AI That Clicks, Types, and Books for You

Most AI still waits for you to type into a box. A new open-source release wants to flip that around: an AI that uses your computer the way you do — looking at the screen, clicking buttons, filling in forms, and finishing web tasks on its own. It's called NorthCUA, it launched on 3 November 2026, and its biggest model just beat trillion-parameter closed AI systems at the benchmarks that measure real desktop work.

If you've read our coverage of DeepSeek's open-source Super-Agent or OpenAI's computer-using GPT-6.1 Terra, NorthCUA belongs in the same family — but with a twist: the whole thing, weights included, is open for anyone to use and build on. Here's what it is, how it was trained, and whether it matters to you.

In This Article

What is NorthCUA?

NorthCUA is an open-source computer-use agent suite — software that turns an AI model into a virtual employee that can operate real desktop applications and websites. Instead of calling an API, the agent sees screenshots, decides where to click, what to type, and how to scroll, completing multi-step tasks the way a human operator would.

The project ships three model sizes, a technical report (posted to Hugging Face Papers), the model weights, the training code, documentation, and the Harness execution framework that runs the agents through long, messy interactions. All three models are built on Qwen3.5, Alibaba's open-weight model family — the same base family we compared in DeepSeek vs Qwen vs Kimi.

The three models: Nano, Mini, Pro

  • NorthCUA-Nano (9B) — the lightweight option, designed to run on modest hardware and for fast, low-cost agents.
  • NorthCUA-Mini (35B) — the middle tier, and the surprise of the release: it beats a closed rival on each major benchmark (more below).
  • NorthCUA-Pro (397B) — the flagship, built for maximum capability on hard desktop and web tasks.

That tiering matters because computer-use agents have a hardware problem: watching a screen and planning clicks every few seconds is expensive. A 9B model that can handle routine tasks — and a 35B model that punches above its weight — could make self-hosted agents practical for far more teams than the frontier APIs currently do.

How it was trained: the Experience Flywheel

The team describes three co-designed pieces behind NorthCUA:

  • The Experience Flywheel — a pipeline that gathers real-world executable tasks rather than synthetic ones. The release collected more than 4 million GUI interaction steps across applications and web interfaces, giving the models experience with the messy reality of real software instead of clean lab examples.
  • The Harness — an execution framework that keeps an agent going through prolonged, multi-step interactions (think a 40-step booking flow, not a single click), with recovery when something goes wrong.
  • Staged learning — the models are first trained with supervised fine-tuning on the collected interactions, then refined with reinforcement learning in live environments.

This training philosophy — learn from real tasks, in real environments, with room to fail and recover — is the same direction the whole industry is moving. It's what makes the benchmark results below interesting: they're measured on verified task sets, not cherry-picked demos.

Benchmarks: how Pro beats closed giants

The release was tested on two respected computer-use benchmarks: OSWorld-Verified (real desktop tasks across operating systems) and WebArena-Verified (real web tasks on live-style sites). The headline results, as reported in the technical report:

  • On OSWorld-Verified (desktop): NorthCUA-Pro outperforms Qwen3.8-Max and Kimi-K3 — both trillion-parameter-class closed systems. NorthCUA-Mini surpasses Kimi-K3.
  • On WebArena-Verified (web): NorthCUA-Pro outperforms both closed rivals. NorthCUA-Mini surpasses Qwen3.8-Max.

Two honest caveats, straight from the coverage: the leadership claims are stated as comparative verdicts rather than headline numeric scores in the abstract, so wait for independent replication before treating them as settled. And benchmarks measure task success, not judgment — a model that clicks accurately can still do the wrong thing confidently. Still, an open 397B model beating closed trillion-parameter systems on desktop automation is a genuine signal that the capability gap is narrowing — echoing the pattern in our Muse Spark Open coverage.

NorthCUA-Ultrafast: the 2.3x speed variant

Alongside the main models, the release includes NorthCUA-Ultrafast — a speed-optimized variant that pairs direct action selection with generation on the pre-RL Nano model. The report credits it with completing web tasks 2.3 times faster on average than the standard inference setup.

Speed is the underrated variable in computer-use agents. A task that takes a human 2 minutes and an agent 20 isn't useful in production, no matter how accurate. A 2.3x speedup on successful web tasks is the difference between a demo and a deployable workflow — worth watching as independent testers get their hands on it.

Pros and cons

Pros:

  • Fully open release — weights, training code, documentation, and Harness framework all ship with it.
  • Flagship Pro beats trillion-parameter closed models on verified desktop and web benchmarks.
  • Three sizes, so teams can trade capability for cost (Nano and Mini are self-hostable).
  • Trained on 4M+ real GUI interaction steps, not synthetic demos.
  • Ultrafast variant tackles the speed problem head-on (2.3x faster).

Cons:

  • No public demo or hosted API announced — this is a build-it-yourself release for now.
  • Benchmark claims are comparative verdicts, not headline numbers; independent verification is pending.
  • Computer-use agents remain risky around credentials and destructive actions — human oversight is still mandatory.
  • Pro at 397B parameters needs serious GPU hardware to self-host.

Who is it for?

Right now, NorthCUA is a researcher and developer release, not a consumer product. It's for teams building production computer-use agents, RPA (robotic process automation) engineers looking at the open-source landscape, and anyone comparing open-weight agent models against closed APIs.

If you run AI coding assistants or build internal automation, the practical question is the build-vs-buy one: does an open 35B or 397B model on your own hardware beat paying per-token to a frontier API for routine desktop work? NorthCUA gives the open side of that argument its strongest evidence yet.

Verdict

NorthCUA is the most complete open-source computer-use agent release we've seen this year: real weights, real training code, real benchmark wins over closed systems, and a speed-optimized variant in the same package. It's not a product you can sign up for today — but as the open counterweight to closed computer-use offerings, it's the one to watch. If the benchmark verdicts hold up under independent testing, 2027's AI agent on your desktop might be running open weights.

Frequently Asked Questions

Is NorthCUA free to use?

It's an open-source release — the weights, training code, documentation, and Harness framework are published for anyone to use. You'll still need your own hardware (or cloud GPUs) to run the models; only Nano (9B) is realistically self-hostable on modest machines.

What does NorthCUA actually do?

It's a computer-use agent: the AI looks at your screen and performs actions — clicking, typing, scrolling, filling forms — to complete tasks in desktop apps and websites, instead of just answering in a chat window.

How does NorthCUA compare to Claude's computer use or OpenAI's agents?

The reported OSWorld-Verified and WebArena-Verified results have NorthCUA-Pro outperforming trillion-parameter closed models including Qwen3.8-Max and Kimi-K3. The claims are comparative verdicts from the project's own technical report — independent replication is still needed for a definitive ranking.

Can I try NorthCUA today?

There's no announced hosted demo or one-click product — this is a build-it-yourself release aimed at developers and researchers. Expect community wrappers and hosted options to appear as people experiment with it.

Sources: technical report details as covered by AI Weekly (report posted to Hugging Face Papers, 3 Nov 2026), Open Source For You (3 Nov 2026), and the November 2026 AI model release tracker. Benchmark leadership claims are the project's own reported results and await independent verification.