NorthCUA: open-source computer-use agent family claims SOTA on OSWorld and WebArena
Powerful_Evening5495 · reddit · 2026-11-03
NorthCUA is an open system + model family (Nano/Mini/Pro) that scales computer-use agents via an "Experience Flywheel": generating verified executable tasks, recording long-horizon decisions, curating trajectories, evolving reusable skills through recursive self-improvement, and closing the loop with joint desktop-web RL after GUI-focused mid-training and SFT. Benchmarks in the blog figure show NorthCUA-Pro at ScreenSpot-Pro 93.3, OSWorld-Verified 88.6, OSWorld-G 91.0, OSWorld 2.0 66.8, WebArena-Verified 73.9 and WebTestBench 60.3, ahead of Qwen3.8-Max, Kimi-K3, DeepSeek-V4.1-Flash and GLM-5.3-Flash. A NorthCUA-Ultrafast variant trades quality for latency via direct action selection. Code: github.com/beipo-ai/NorthCUA; paper: arXiv 2611.02102.
Related event: Shanghai Qiji Zhifeng Open-Sources NorthCUA Computer-Use Agent(2 posts)→
More from coding & agent
- Article explores how AI agents can improve their own harness without retraining — dima806_dima · 2026-11-03
- Slop Cannon plugin for Nous Research's Hermes agent renders multi-scene videos with a native Studio UI — Teknium · 2026-11-03
- Honest review of esolang Bend2: no type inference, 9/10, 'still too easy for JS devs' — max_paperclips · 2026-11-03
- Dev claims WebMCP beats browser-use agents for web automation — aniketmaurya · 2026-11-03
- CUA-Master: Nested Dual-Loop Framework Scales Computer-Use Agents via Open-World Recursive Improvements — Xinlei Yu · 2026-11-03
- CrossSWE benchmark shows coding agents struggle with cross-repository development — Silin Chen · 2026-11-03