# NorthCUA Models and Harness: Benchmarks, Downloads and What Is Actually Available

NorthCUA pairs open computer-use models with a reusable runtime, but its downloadable checkpoint and reported scores need to be read separately.

By Paras Madan | 2026-11-03

NorthCUA is a computer-use model family and an open Desktop/Web agent harness from Shanghai Qiji Zhifeng. Its November 1 technical report describes training agents to finish software tasks, while updated November 3 documentation confirms public model resources. The useful distinction is what developers can download versus what the research evaluated.

## NorthCUA Nano, Mini and Pro

The [NorthCUA technical report](https://arxiv.org/abs/2611.02102) describes three Qwen3.5-based sizes: Nano at 9 billion parameters, Mini at 35 billion total and 3 billion active, and Pro at 397 billion total and 17 billion active.

In the authors' OSWorld-Verified evaluation, Nano reaches 78.9%, Mini 85.3% and Pro 88.6% task success. That evaluation uses 361 tasks with a 100-step limit. These are reported research results, rather than independent measurements of the current public download.

The more demanding OSWorld 2.0 result gives a useful counterweight. Pro records 30.6% complete-task success and 66.8% partial credit across 108 tasks. Making progress and finishing every requirement remain very different outcomes.

## What can developers download?

The [maintained NorthCUA repository](https://github.com/beipo-ai/NorthCUA) provides an Apache-2.0 harness built on NexAU. It combines visual context, working notes, outcome checks and recoverable batches. The earlier NorthCUA-Harness repository is now a migration notice pointing to this maintained location.

The harness is model-agnostic. Developers can connect a vision model with structured tool calls, so evaluating the runtime does not require loading Pro. Its local no-key tutorial checks integration with a deterministic model; passing that tutorial does not demonstrate autonomous task-solving ability.

The project's [model and endpoint documentation](https://github.com/beipo-ai/NorthCUA/blob/main/docs/models.md) records public Nano, Mini and Pro model resources under Apache-2.0. It also says its November 3 availability check supersedes an earlier HTTP 401 observation for Pro. This is a concrete improvement for anyone who previously encountered an inaccessible checkpoint.

## The Pro checkpoint has an important benchmark caveat

The [NorthCUA-Pro model card](https://huggingface.co/china-beipo/NorthCUA-Pro) identifies the export as post-training checkpoint 975 and explicitly says the report's research scores have not been tied to that exact export. **Available weights do not yet establish benchmark reproduction.**

Pro's files occupy approximately 793.63 GB before runtime caches, across 122 weight shards. Its 17 billion active parameters describe computation through the mixture-of-experts model, not the memory needed to hold all the weights. That difference matters when estimating whether a deployment is practical.

The card gives a configured context length of 262,144 tokens and an illustrative loading example. It does not present that example as a benchmarked serving configuration. Teams still need to check image handling, tool calls and memory overhead with their chosen engine.

## Why the harness is worth evaluating separately

NorthCUA's research combines verified task trajectories, long-horizon supervision and desktop–web reinforcement learning. Hard tasks feed reusable skills and verified intermediate milestones back into later attempts.

For a product team, the practical question is whether its own workflows benefit from that continuity. A browser agent that remembers why it rejected an earlier option may avoid repeating work. An agent that checks the final artifact can catch a different class of failure from one that only checks whether each click succeeded.

A sensible evaluation separates three questions: can the model choose useful actions, can the runtime preserve the task over a long session, and does the final output meet every requirement? NorthCUA supplies pieces to investigate all three. The next meaningful result will be a reproducible connection between a named checkpoint, its harness settings and complete task outcomes.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn AI launches and engineering work into clear stories that developers and founders read and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).