The NorthCUA project published its model family on Hugging Face under beipo-ai: NorthCUA-Nano (9B, dense), NorthCUA-Mini (35B total, 3B active) and NorthCUA-Pro (397B total, 17B active), all Apache 2.0. Each takes a request, screenshots and interaction history, and returns reasoning, replies and tool calls for a desktop or browser runtime. It's independent post-training on Qwen3.5, not an official Qwen release. According to AITrending's read of the upload history, Nano's weights finished uploading on November 2.
The model doesn't click anything by itself. The card says to run it with the official NorthCUA Harness on GitHub, which supplies the tools, runs the actions and sends back new screenshots. Coordinates are a strict contract: x and y sit in a 0 to 1000 space relative to the screenshot, and the harness turns them into pixels.
The report's numbers, all self-reported: on OSWorld-Verified (361 tasks, 100-step limit) Nano scores 78.90%, Mini 85.30% and Pro 88.64%. On WebArena-Verified the three get 67.16%, 70.43% and 73.89%. The card says plainly that these came from research checkpoints and aren't new measurements of the exported files. Training used more than 100,000 GUI trajectories and 4 million interaction steps in mid-training, then about 3,000 verified long trajectories.
Nano's files come to about 18.84 GB, mostly BF16 weights, per its card as summarized by AITrending.