Cua Releases CUA-S1 Computer-Use Models on GitHub — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Cua released CUA-S1, a family of small, specialized System 1 models for computer-use decisions focused on fast, bounded actions like scoring which value belongs in a form field rather than generating full responses token by token.
- CUA-S1's first research profile targets forms — scoring decisions from structured interface elements while application code orders the actions, with execution optionally handled through Cua Driver and explicit action boundaries.
- The Cua project ships as MIT-licensed code on GitHub, with model weights hosted separately on Hugging Face; the CUA-S1-FORMS model and its dataset card specify scope, limitations, and artifact-specific licenses.
- Cua Fleets provision isolated Linux cloud desktops via the Sandbox SDK, letting agents claim a desktop from a pool, run commands, capture screenshots, and clean up afterward.
- Cua Driver connects agents — including Claude Code, Codex, Cursor, and OpenClaw — to native desktop apps and browsers on macOS, Windows, and Linux through CLI, MCP, or typed SDKs, with background delivery that avoids moving the user's pointer where platforms allow.
- Lume creates and manages local macOS and Linux VMs on Apple Silicon via Apple's Virtualization.Framework, with a CLI that can spin up a vanilla macOS Tahoe VM from an Apple restore image and expose it over SSH.
- Cua Bench lets researchers build computer-use tasks, evaluate agents, and export trajectories for training, with simulated tasks that require no VM, Docker, or model API key and support for Python 3.12 or 3.13 with uv.
Why it matters: Cua bundles the full computer-use stack — models, cloud and local VMs, native drivers, and benchmarks — under one MIT-licensed project, so researchers and developers can plug in their own agent and skip stitching together separate tools. The System 1 framing positions CUA-S1 for narrow, bounded interface decisions rather than competing with the general-purpose planners (Claude Code, Codex, Cursor) the same project also integrates with.
Ask SkimNews
