# NorKa: The Open-Weight Model Family Built for Computer Use > Source: > Published: 2026-11-06 00:00:00+00:00 # NorKa: The Open-Weight Model Family Built for Computer Use NorKa's Pro, Mini, and Nano models are open-weight computer-use agents, with the Pro variant hitting 88 on OSWorld Verified. ## What is NorKa? NorKa is an open-weight model family built specifically for computer use, meaning it can look at a screen, decide what to click or type, and carry out multi-step tasks on a desktop the way a human operator would. It ships in three sizes (Pro, Mini, and Nano) and comes paired with its own open-source harness from Beo AI. On OSWorld Verified, a widely used benchmark for computer-control agents, the Pro variant scores 88, putting it in frontier territory alongside closed models from the major labs. ## TL;DR - **NorKa** is a family of open-weight models (Pro, Mini, Nano) purpose-built for controlling a desktop environment rather than just generating text. - The **Pro model** is a mixture-of-experts design with 397 billion total parameters and 17 billion active parameters, and it scores 88 on OSWorld Verified. - The **Nano model** has only 9 billion total parameters yet scores 78.9 on the same benchmark, a gap of just about 10 points from the much larger Pro. - NorKa ships with its own **open-source harness** (from Beo AI) that handles the screenshot-observe-think-act loop needed for desktop control. - Benchmark results on **WebArena Verified** show a similarly tight spread between Nano and Pro, suggesting the smaller models capture most of the capability at a fraction of the compute. - Because weights and harness are both open, developers can self-host a **computer-use agent** instead of relying on a closed API, with the Nano size plausible to run on modest hardware. - NorKa arrived the same week as several other major open-weight releases, including Kimi K3.1 Pro, reflecting a broader push toward frontier capability outside the closed labs. *made for the enterprise.* **↑** *Approved* *Approved* *Approved* *Approved* *Approved* *Approved* *Approved* *Approved* *Approved* *Approved* ## How does NorKa actually control a computer? NorKa’s harness runs a loop that mirrors how a human would operate a desktop: observe, think, act. The agent takes a screenshot of the current screen state, reasons about what it’s looking at and what step comes next, then issues an action such as a click, a keystroke, or a cursor move. That action changes the screen, a new screenshot is taken, and the loop repeats. This observe-think-act cycle is the standard architecture behind most computer-use agents, and it’s the same basic pattern used by closed systems from the major AI labs. What makes NorKa notable is that both the underlying model and the harness that runs this loop are open, so the whole pipeline can be inspected, modified, or self-hosted rather than accessed only through a vendor’s API. ## What do the benchmark scores actually mean? Two benchmarks matter most for evaluating a computer-use model: OSWorld Verified and WebArena Verified. OSWorld tests an agent’s ability to complete real tasks inside an actual operating system environment, things like file management, application switching, and multi-step workflows. WebArena tests the same kind of task completion but scoped to web-based environments. Both are considered legitimate, hard-to-game measures of whether an agent can actually get things done, as opposed to benchmarks that just test static question-answering. On OSWorld Verified, NorKa Nano (9 billion total parameters) scores 78.9, NorKa Mini (35 billion total parameters) scores 85, and NorKa Pro (397 billion total parameters, 17 billion active) scores 88. The scores on WebArena Verified follow a similarly tight pattern across the three sizes. What stands out is the small gap, about 10 points, between the smallest and largest models despite a roughly 44x difference in total parameter count. That’s unusual: in most model families, score curves between a 9B model and a 397B model are much steeper. ## Why does the size-to-performance ratio matter? The practical implication is that a lot of computer-use capability doesn’t require a massive model. NorKa Nano, at 9 billion parameters, is small enough to plausibly run on consumer hardware, and its creators have suggested it could even run on a phone-class device. If a 9B model can clear nearly 79 on OSWorld Verified while the 397B flagship clears 88, then for many real-world automation tasks the smaller model may be the more practical choice: cheaper to host, faster to run, and easier to deploy at the edge, without giving up a large share of task-completion accuracy. This matters for anyone building agentic tooling around desktop or browser automation. A model that needs a datacenter-grade GPU cluster to run is a non-starter for a lot of use cases like local automation tools, on-device assistants, or privacy-sensitive workflows where sending screenshots to a third-party API isn’t acceptable. An open, smaller model that gets most of the way to frontier performance changes the calculus. ## How does NorKa compare to other computer-use models? *mission-critical*enterprise applications. **Dr. A. Okafor** Cardiology · St. Mercy Hospital - Patient records*✓ Granted* - Surgery schedule*✓ Granted* - Billing export*✕ Denied* NorKa’s Pro score of 88 on OSWorld Verified puts it in the same range as the strongest models currently available, including some closed frontier systems. For context, Kimi K3.1 Pro, a 3 trillion parameter open-weight model also released the same week, scored third place on OSWorld, behind a couple of leading closed models. NorKa isn’t a general-purpose chat or reasoning model the way Kimi K3.1 is; it’s narrowly built and tuned for the specific task of operating a computer interface. That specialization appears to pay off: a model purpose-built for computer use, even at a fraction of the parameter count of a general frontier model, can match or exceed general models on computer-use-specific benchmarks. This is part of a broader pattern in open-weight releases: rather than one model trying to do everything, labs are increasingly shipping task-specialized models (computer use, scientific discovery, decision-making, voice-to-action) that outperform generalist models on their specific domain while using far less compute. ## Is NorKa worth setting up? For developers specifically interested in agentic desktop or browser automation, NorKa is worth evaluating because it’s fully open, the harness is open-source, and the benchmark numbers are competitive with frontier systems. Setup involves pulling the model weights from Hugging Face and running them through the accompanying harness, which handles the screenshot-loop logic. Given that the Nano variant is small enough to be reasonably accessible on local hardware, teams can experiment without committing to expensive GPU rental. The tradeoff is that computer-use models, even strong ones, aren’t foolproof. An 88 on OSWorld Verified means the model fails roughly 12% of tasks in a controlled benchmark, and real-world environments tend to be messier than benchmark environments. Anyone deploying this kind of agent in production should expect to build in human review or fallback logic for failed or ambiguous actions, especially for anything touching files, payments, or irreversible actions. ## Frequently Asked Questions ### What is NorKa’s biggest model and how large is it? NorKa Pro is a mixture-of-experts model with 397 billion total parameters and 17 billion active parameters per forward pass, which keeps inference costs lower than a dense model of the same total size. ### What benchmark score does NorKa get on OSWorld Verified? NorKa Pro scores 88 on OSWorld Verified. NorKa Mini scores 85, and NorKa Nano, at just 9 billion parameters, scores 78.9. ### Can I run NorKa locally? The Nano variant, at 9 billion total parameters, is small enough to run on consumer-grade hardware or potentially even mobile devices. The Pro variant, at 397 billion total parameters, requires significantly more compute, typically multiple high-memory GPUs or an inference provider. ### What is the NorKa harness? It’s an open-source runtime, built by Beo AI, that manages the observe-think-act loop a computer-use agent needs: taking screenshots, reasoning about the next step, and executing actions like clicks or keystrokes. ### How is NorKa different from a general-purpose language model? General-purpose models like Kimi K3.1 are built for broad reasoning, coding, and conversation. NorKa is narrowly trained for operating graphical interfaces, which is why its smaller variants can match or beat much larger general models specifically on computer-use benchmarks.