Cua has released CUA-S1, a family of small specialist models for computer-use decisions, alongside an open-source stack for desktop automation, isolated cloud desktops and computer-use benchmarks. The pitch is simple: not every step of a computer-use task needs a frontier model, and treating every click that way is what makes desktop agents slow and expensive. The repository describes the project as tooling to scale “computer-use 2.0” with open-source drivers, cross-OS fleets and evaluation benchmarks.
Key facts
| Fact | Detail |
|---|---|
| Question answered | Computer-Use 2.0: Specialist Models for the Small Decisions |
| Short answer | Cua shipped CUA-S1, a family of small specialist models for computer-use decisions, plus open-source drivers, cloud desktops and benchmarks. |
| Reading time | ~2 min |
| Last updated | 2026-09-20 |
The split they are arguing for
On Hacker News the team behind it framed the problem as two different kinds of work inside one task:
- Some steps are planning problems - think through a route, explore alternatives, recover from a failure.
- Other steps are local decisions - this value belongs in that box, this checkbox should be ticked, this element can be ignored.
CUA-S1 targets the second kind: small, specialised models that answer fast, while a larger model handles the planning when a task genuinely needs it. The same repository ships a broader stack around that idea - open-source drivers, isolated cloud desktops (“fleets”), local macOS VMs and a benchmark suite for evaluating computer-use agents.
Why the framing is worth reading
The repository’s own definition of Computer-Use 2.0 is useful: an agent moving between code, APIs and graphical interfaces within the same task.
That corrects a common mental model. A desktop agent is not a robot clicking around a screen; it is a system that decides, step by step, whether the cheapest correct route is a shell command, an API call, or a pointer on a UI element. Once you think of it that way, splitting decisions by difficulty stops being an optimisation detail and becomes the thing that decides whether a long task is affordable.
What this means if you run an agent on your own machine
On your own PC the split that matters is not by model size, it is by consequence:
- Low-stakes, reversible steps - reading a folder, drafting a plan - can run without ceremony.
- Steps that write, delete, send or spend should stop and ask, whichever model proposed them.
That is the design MOOGH follows on Windows: it plans first, shows the proposed changes, executes only after you approve, and keeps an action history you can read back. A faster model underneath changes how quickly the work happens; it does not change who is accountable for it. Download MOOGH if you want the approval step to stay with you.
Related reading: