NorthCUA: an open model family trained to operate desktop and web apps

A team publishing as Beipo has released NorthCUA, a family of open models trained to use a computer the way a person does: look at the screen, decide, then click, type or scroll. The models and their technical report appeared on 1 November 2026, with weights on Hugging Face under the Apache 2.0 licence.

Three sizes, one runtime

• NorthCUA-Pro: 397 billion parameters, 17 billion active (mixture of experts).
• NorthCUA-Mini: 35 billion parameters, 3 billion active.
• NorthCUA-Nano: a dense model with 9 billion parameters.

All three are built on Qwen3.5. They run inside the NorthCUA Harness, an open-source runtime that prepares a desktop or browser environment, carries out the agent’s actions, checks the outcome and keeps a record of each run. The model receives the request, screenshots and the interaction history, and returns its reasoning plus tool calls. Screen coordinates are given on a 0 to 1,000 grid, and the harness converts them to pixels.

Reported results

On OSWorld-Verified, a benchmark of 361 desktop tasks with a 100-step limit, the team reports success rates of 88.64% for Pro, 85.30% for Mini and 78.90% for Nano. On WebArena-Verified, 812 web tasks, the figures are 73.89%, 70.43% and 67.16%. The gap between the largest and smallest models is about ten points, which matters because the 9B model needs a fraction of the hardware. The project page lists caveats on how some numbers were collected, so read them as the team’s own measurements.

How it was trained

The report describes an ‘Experience Flywheel’ that turns goals into executable tasks with automatic checks, then feeds verified attempts back into training. That happened in three stages: broad interface training on more than 100,000 trajectories, supervised fine-tuning on about 3,000 verified long tasks, and reinforcement learning across desktop and web tasks together. A faster variant, NorthCUA-Ultrafast, chooses from a fixed set of actions for low-latency web work.