Northka Computer Use Model: Which Size and How Good?

How good is the Northka computer-use model?

Northka is an open-weight model family built to control a computer. It comes in three sizes: Pro (397B total parameters, 17B active, mixture of experts), Mini (35B total) and Nano (9B total). On OSWorld Verified, a computer-control benchmark, Nano scores 78.9, Mini scores 85 and Pro scores 88. The gap between the smallest and largest model is only about nine points, so the small models are a serious option.

What are the Northka Pro, Mini and Nano models?

Northka is a computer-use model family. Its job is to look at a screen, decide what to do, and operate the interface the way a person would. The weights are open, and so is the harness that runs it.

Pro is a mixture-of-experts model. That means only 17B of its 397B parameters activate for a given token, which keeps inference cost closer to a mid-sized dense model than the headline count suggests. You still need memory for all the weights, though, so plan hardware around 397B, not 17B.

The source does not give architecture details for Mini or Nano, so I won't guess at them.

How do the sizes score on OSWorld Verified and WebArena Verified?

OSWorld Verified is the number to anchor on. It tests an agent on real desktop tasks across applications, and the "Verified" label indicates a cleaned-up version of the benchmark.

• Nano (9B): 78.9
• Mini (35B): 85
• Pro (397B): 88

Going from Nano to Mini buys about six points. Going from Mini to Pro buys three more. Returns drop off fast as the model grows.

WebArena Verified, which measures browser-based tasks, tells a similar story. The scores for Nano and Pro are described as very close, and all are strong frontier-grade results. I don't have exact WebArena numbers from the source, so check the model page before quoting them.

For context, a different open-weight release covered the same week, Kimi K3.1 Pro, was reported as third on OSWorld behind two closed frontier models. Northka's 88 is not a like-for-like comparison with that, since I have no matching score for it, but it shows computer use is a category where open weights are competing near the top.

Which Northka size should you use?

Pick by hardware first, then by how much failure you can tolerate.

• Nano (9B) fits most hardware. The source even suggests it could run on a phone, though that is speculation, not a tested claim. It is the obvious starting point for local experiments, prototypes and cost-sensitive automation.
• Mini (35B) is the middle option. It lands at 85 on OSWorld Verified, three points under Pro, and needs far less memory than Pro.
• Pro (397B, 17B active) gives the best score at 88. It makes sense when each failed task is expensive and you have the infrastructure to host a model of that size or a provider that serves it.

A ten-point benchmark gap does not translate evenly to your workflow. Long, multi-step tasks compound errors, so a few points of per-task accuracy can matter more in practice than the average suggests. If your tasks are short and well bounded, Nano may be enough. If they chain dozens of actions, test Mini and Pro on your own workload before committing.

How does the Northka harness work?

The model runs inside a custom open-source harness from Beo AI. For the desktop agent, the harness runs a repeating loop:

1. Screenshot. The harness captures the current screen.
2. Observe and think. The model reads what is on screen and reasons about the next step toward the goal.
3. Act. The model issues an action, such as a click or keystroke, which the harness executes.
4. The loop repeats with a fresh screenshot until the task is done.

This is the standard pattern for vision-based computer use. The model never gets an accessibility tree or an API to the app. It sees pixels and acts on them, which is why it can work across arbitrary software. It also means latency adds up, since every step costs a screenshot and a model call. A smaller model like Nano can make that loop considerably cheaper per step.

How do you set up Northka?

The Hugging Face model page has the setup instructions and links to the Beo AI harness. The creator who demonstrated it noted that the process looks simple enough that you can hand the page link to a coding agent and ask it to set the project up for you. That is a convenience, not a requirement, and you should still review what an agent installs and runs, especially for a tool that controls your desktop.

Practical precautions for any computer-use agent:

• Run it in a virtual machine or a disposable container first.
• Keep it away from accounts and files you cannot afford to lose.
• Log screenshots and actions so you can debug failed runs.
• Start with a short, bounded task before attempting long workflows.

Is Northka worth trying?

For anyone building agents, yes. Scores of 78.9 to 88 on OSWorld Verified from open weights, with an open harness included, give you a computer-use stack you can inspect, self-host and modify. The Nano result is the most interesting part. A 9B model within about nine points of the largest model changes the cost math for running agents at scale.

The caveats are real. Benchmarks are not your workflow. The source gives no WebArena numbers, no hardware requirements for Pro or Mini, and no latency figures. Treat the scores as a reason to run your own evaluation, not as a guarantee.