NorthCUA: open computer-use agents built on Qwen3.5, with a 397B Pro that beats trillion-parameter closed models on desktop benchmarks
Three open-source computer-use agents built on Qwen3.5 — Nano at 9B, Mini at 35B, and Pro at 397B — plus the harness code, with Pro reportedly outperforming trillion-parameter closed models Qwen3.8-Max and Kimi-K3 on OSWorld-Verified and WebArena-Verified, and an Ultrafast variant running 2.3x faster on matched WebArena tasks.
Why it matters
It is the first open computer-use family to claim wins over trillion-parameter closed rivals on standard desktop and web benchmarks, with a small 9B model in the same release for on-device use.
What you could build with it
A QA-automation startup could run NorthCUA Mini on its own machines to replay web-app regression suites by watching the screen and clicking through flows, replacing brittle selector-based scripts with agents that adapt when the UI changes.
Does it hold up?
Too early to judge: benchmark claims come from the authors' technical report with verdicts rather than numeric scores in the abstract, and no public weights or demo link were verified at ingest time.
Built with NorthCUA
- NorthCUA Pro 397B tops Qwen3.8-Max and Kimi-K3 on OSWorld-Verifiedarticle · Breakdown of the three model sizes, the Experience Flywheel and harness design, and the benchmark verdicts against closed rivals.
Learn more
First spotted on article: source.
More AI agents
CowAgent: open-source super AI assistant and agent harness with memory that compoundsAn open-source personal AI assistant and agent harness by zhayujie, back on Hacker News' front page this week: it…agents · JEV 0.78Agent Desktop: Vercel Labs' CLI that gives agents reliable desktop control via the accessibility treeVercel Labs' Rust CLI and C library lets AI agents read and operate desktop apps through OS accessibility trees: a…agents · JEV 0.76Open Notebook: self-hosted, multi-model NotebookLM alternative trending againOpen-source research notebook that turns documents, web pages, audio and video into searchable notebooks with chat…agents · JEV 0.75OpenBot: self-hosted AI coworkers, each with its own computer and audit trailCopilotKit's MIT-licensed alpha gives each AI coworker its own container with a browser, files and logins, behind a…agents · JEV 0.73DeepTutor: open-source, agent-native personalized tutor trending againHKUDS/DeepTutor, the University of Hong Kong's open-source personalized learning assistant, is trending on GitHub…agents · JEV 0.72claude-seo: universal SEO skill for Claude Code, trending on GitHubAgrici Daniel's MIT-licensed Claude Code plugin bundles 26 sub-skills and 19 specialist agents that run full SEO…agents · JEV 0.72
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.