NorthCUA: The Open-Source AI That Uses Your Computer for You
Most AI waits for you to type into a box. NorthCUA, released 3 November 2026, flips that around: it's an open-source computer-use agent suite that sees your screen, clicks buttons, fills forms, and finishes web tasks on
Most AI waits for you to type into a box. NorthCUA, released 3 November 2026, flips that around: it's an open-source computer-use agent suite that sees your screen, clicks buttons, fills forms, and finishes web tasks on its own. And its flagship model just beat trillion-parameter closed systems on the benchmarks that measure real desktop work.
I wrote the full breakdown on NeurAITool β here's the developer-relevant core.
The three models
Built on Qwen3.5, three sizes ship together:
- Nano (9B) β lightweight, for modest hardware and fast cheap agents
- Mini (35B) β the surprise: beats a closed rival on each major benchmark
- Pro (397B) β the flagship for maximum capability
The whole release β weights, training code, docs, and the Harness execution framework β is open.
The training recipe
Three co-designed pieces: an Experience Flywheel that gathered 4M+ real-world GUI interaction steps across apps and web interfaces, a Harness that keeps agents going through long multi-step interactions with recovery, and staged learning (supervised fine-tuning, then reinforcement learning in live environments).
Benchmarks that matter
On OSWorld-Verified (desktop) and WebArena-Verified (web), NorthCUA-Pro outperforms trillion-parameter closed models Qwen3.8-Max and Kimi-K3. Mini surpasses Kimi-K3 on desktop and Qwen3.8-Max on web. Caveat: these are the project's own reported comparative verdicts β independent replication is still pending.
There's also a NorthCUA-Ultrafast variant: 2.3x faster on matched successful WebArena tasks. Speed is the underrated variable in computer-use agents β this is what makes the difference between a demo and a deployable workflow.
The build-vs-buy question
No hosted demo or API is announced yet β this is a build-it-yourself release for developers and researchers. The practical question for teams: does an open 35B or 397B model on your own hardware beat paying per-token to a frontier API for routine desktop work? NorthCUA is the strongest evidence yet for the open side of that argument.
Full analysis, pros/cons, who it's for, and FAQ on the original post: https://www.neuraitool.com/2026/11/northcua-open-source-ai-that-uses-your.html
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.