A 397B Open Model Now Drives Your Desktop. The 9B One Is Only Nine Points Behind.

NorthCUA shipped on November 3 as a set of three open-weight models and the harness that runs them. The headline number is Pro, a 397 billion parameter mixture-of-experts system that scores 88 on OSWorld-Verified, a benchmark where an agent has to complete multi-step tasks inside a real operating system: moving files, switching applications, filling forms, reading a screen and deciding what to click next.
The number that should get more attention is 78.9. That is Nano, at 9 billion parameters.
Nine points of benchmark gap across a 44-fold difference in model size. Mini sits in the middle at 35 billion and 85 points. On WebArena-Verified, the browser-navigation benchmark, the spread is similarly narrow.
What the benchmark actually measures
OSWorld-Verified gives an agent a task and a virtual machine and grades whether the end state matches. It is closer to real work than a multiple-choice test, and further from real work than it looks. The tasks are well specified, the environment is clean, and nothing unexpected interrupts.
A score of 88 means roughly 12 percent of tasks fail. That is a good result for a research system and a bad result for a product that touches your files. If an agent gets one task in eight wrong, and some of those failures involve deleting the wrong thing or clicking the wrong button, nobody is going to leave it unattended on a work laptop.
That is the honest reading of the release, and to the authors' credit the technical report does not oversell it. It also does not publish absolute scores for every comparison, framing the closed-model results as verdicts rather than tables: Pro beats Qwen3.8-Max and Kimi-K3 on OSWorld, Mini beats Kimi-K3, and on WebArena Mini beats Qwen3.8-Max while Pro beats both.
Those are the authors' own evaluations on their own task set. Treat them as a claim, not a leaderboard position.
How it was trained
Three pieces, and the naming is unusually literal.
The Experience Flywheel supplies executable tasks drawn from real application usage. The point is that the tasks come with ground truth, because you can check whether the spreadsheet got filled correctly.
The Harness holds a long interaction together: it records decisions, feeds observations back, and validates outcomes across sessions that run far longer than a single context window. Every trajectory was collected in an isolated E2B sandbox, which also means the training data is reproducible and the model never saw a live machine it could break.
Staged Learning is the schedule. It starts with grounding mid-training, teaching the model what a button and a text field look like. Then long-horizon supervised fine-tuning on complete task trajectories. Then joint reinforcement learning across desktop and web using Group-Relative Policy Optimization, the same family of method behind most of the recent reasoning-model gains.
The scale is 4 million plus GUI interaction steps across dozens of applications. That is large for GUI data and small compared to the text corpora behind a frontier language model, which is exactly the condition under which specialization wins.
The Ultrafast variant is the more interesting artifact
NorthCUA-Ultrafast pairs direct action selection with generation, running on the pre-RL Nano model. In plain terms, it skips the reasoning step and maps observation to action through learned shortcuts, falling back to thinking when the shortcut is not confident.
The report credits it with 2.3 times faster average completion on matched successful WebArena tasks. Same tasks, same success, less than half the wall clock.
That is the shape of the deployment problem for computer-use agents, and it is not the shape most people expect. Success rate gets the attention. Latency is what determines whether a human will actually watch the agent work or leave the room. A task that takes four minutes and succeeds 85 percent of the time is a tool. The same task at 100 seconds is a feature.
It also implies something about where the intelligence has to live. If a 9B model with shortcuts is nearly as good as a 397B model thinking hard, then for a large class of desktop tasks the bottleneck was never reasoning depth. It was grounding quality and data coverage, and those are cheaper to improve.
Why open weights matter here more than elsewhere
A computer-use agent sees your screen. Every screenshot it captures can contain a customer record, a password field, an unannounced product, an internal dashboard. Sending that stream to a remote API is a data governance decision. For plenty of organizations it is a decision they are not allowed to make.
That governance constraint, rather than raw capability, is what this release is selling into. Hospitals, banks, government contractors and manufacturers that want desktop automation and cannot route screen captures through a third party now have three sizes to choose from, plus the harness, plus no software license fee.
The practical guidance that follows is boring and correct. Start with Nano on the three tasks you do most often: filling a form, exporting a report, moving data between two applications. If it can do those, Mini handles production volume on a single data center GPU. Pro exists to find the ceiling, and if your ceiling problem is not solved by Pro, the problem is your task design rather than the model.
Three sizes is a pricing decision
Shipping 9B, 35B and 397B in one release is a way of letting customers sort themselves.

Nine billion parameters fits on a laptop or a mini PC. That is the version a small business in a country with expensive cloud egress can run, and it is the version that makes the release relevant outside the usual few markets. Thirty-five billion fits on one data center GPU, which is the tier most production deployments will land on. Three hundred ninety-seven billion exists to establish that the architecture scales, and to be quoted in the abstract.
The arXiv-style report does something similar with its evaluation. It compares against trillion-parameter closed models rather than against open peers, because the comparison that gets cited is the one where you win. A reader who wants to know how Nano does against a well-tuned 32B vision-language model will not find it here.
None of that makes the numbers wrong. It makes them narrow, and the useful move for a buyer is to ignore the headline and test the tier you can afford.
The part nobody has solved
Twelve percent failure is a statistic. Behind it are specific behaviors that all computer-use systems share.
They misread state. An agent clicks a checkbox that was already checked, and the next step assumes it just turned it on. They lose track of which window has focus. They retry an action that partially succeeded and double the effect. And they have no concept of irreversibility, so the same confidence applies to opening a menu and to pressing send on an email to four hundred people.
Nothing in this release addresses that. The mitigation is architectural: run the agent in a container, snapshot before each irreversible step, require human confirmation on a defined list of actions, and log every click with a frame capture so a failure can be diagnosed.
Open weights make that architecture cheaper to build, which is the real argument for them. A closed API forces you to trust the vendor's safety layer. An open checkpoint lets you put your own fence around it.
The nine points between 9B and 397B say something about the next two years of this market. The models will get better slowly, and the products will be built around whichever tier is good enough at a cost that fits. If Nano is good enough for your workflow today, the interesting engineering sits in the harness you build around the model you already have.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.