NorthCUA Pro 397B Tops Qwen3.8-Max, Kimi-K3 on OSWorld-Verified

NorthCUA Pro, an open-source 397-billion-parameter computer-use agent built on Qwen3.5, outperforms trillion-parameter closed models Qwen3.8-Max and Kimi-K3 on the OSWorld-Verified desktop benchmark and the WebArena-Verified web benchmark, according to a technical report posted to Hugging Face Papers.

The authors ship three model sizes, Nano at 9B, Mini at 35B, and Pro at 397B, plus the Harness code that drives them. The report pitches three co-designed pieces: an Experience Flywheel that supplies executable tasks, a Harness that it says "Sustains long interactions, records decisions, and checks outcomes," and a Staged Learning pipeline that moves from grounding mid-training through long-horizon supervised fine-tuning to joint desktop and web reinforcement learning using Group-Relative Policy Optimization.

Training scale is "4M+ GUI interaction steps across dozens of applications," with trajectories collected in isolated E2B sandboxes and filtered by outcome. A tradeoff variant called NorthCUA-Ultrafast pairs direct action selection with generation on the pre-RL Nano, and the report credits it with "2.3× faster average completion of matched successful WebArena-Verified tasks."

The leadership claims are stated as comparative verdicts rather than numeric scores in the abstract. On OSWorld-Verified, "Pro outperforms Qwen3.8-Max and Kimi-K3; Mini surpasses Kimi-K3." On WebArena-Verified, "Mini surpasses Qwen3.8-Max; Pro outperforms both."

The release lands into a dense run of open-source agent work on our radar; our agents tracker counts 527 stories in the last 90 days, with the Tsinghua and Berkeley Uni-Robot-Agent harness crossing the wire the same day.