1/ Today we're launching Cua Driver Bench, together with Scale AI's AgentEnv. Can an agent do a desktop task in the background while you keep working undisturbed? On macOS, Cua Driver fully passed 58%, Codex computer use 33%, Vercel's agent-desktop 20%, browser-use 3%.

1/ Introducing the Cua Spaces SDK. Every big lab now ships a bot with its own computer: Grok Bot, Muse, dots. None let you bring your own. With Cua Spaces, now you can. Put Claude Code, Codex or any agent on compute you choose, and run jobs across all your computers at once.

1/ We're reimagining your computers as shared spaces for your team and its agents. Today we're excited to share Cua Spaces Teams and Pro, built on Cua Driver, our virtualization stack and remote desktop protocol. On macOS, Linux and Windows. Start free:

1/ We're reimagining what it means for your agents to work with all your computers. Today we're excited to share Cua Spaces, built on Cua Driver and our virtualization stack, rolling out on macOS today. Download it now, it's free and source-available:

1/ Today we're announcing the stable Cua Driver release for Omarchy - a new foundation for computer use, built into the OS from the ground up. Over the last month, we worked directly with @dhh, @SpencerGBull and @vaxryy to bring a native synthetic cursor to Omarchy's Hyprland compositor, enabling true multi-cursor computer use at the OS level. Cua Driver is open source and available at

Cua just released Cua-S1-4B-0.2 today. Here's what you need to know. The Cua team, makers of the computer-use driver behind Hermes Agent, has open-sourced this multimodal decision model under Apache-2.0. It's described as the first multimodal decision model trained with RLOO (Reward-Weighted Leave-One-Out) on live computer-use tasks, using task-completion rewards. The model runs on a frozen Qwen3.5-4B base with LoRA adapters. At each step it takes in the screen state, the task goal, and a fixed set of candidate actions, then returns one action, no accessibility tree required. Training happened in two stages: supervised fine-tuning to teach the decision format, then agentic RL in live cua-bench-basic environments where only a fully completed task earns reward. On the GUI-360 benchmark, it scored 92.9% versus 60.1% for comparison. Text and multimodal adapters, training code, and benchmark results are all merged into the Cua GitHub repo and live on Hugging Face. This is version 0.2, up from 0.1 released the day before. Key numbers: - 4B parameters (Qwen3.5-4B base) - 92.9% vs 60.1% on GUI-360 benchmark - Apache-2.0 license - v0.1 to v0.2 in one day The model is small enough to run on-prem, and developers are already testing it on local hardware for business workflows.

1/ You can now try Cua-S1-4B-0.2 in your browser. Thanks to @multimodalart at @huggingface for building the Space! Pick an example and see how the model scores the candidate actions.

Awesome innovation from the @trycua team! We dont need to burn big model inference just to decide what element of a UI to interact with. Using a local small specialized model to handle execution is the way. Cua-S1 is applying almost the same philosophy as Jev, one layer further down. • screen + candidate actions ↓ S1 ↓ • element • action • abstain

Love the work cua is doing esp with this new release. If you haven't tried computer use models in 6 months worth trying again because they are really good. And if cua does a really good job, then ALL software with various interfaces get abstraced and we can expect massive productivity gains across the economy as a ton of software like EHRs, legacy windows desktop apps, etc become newly accessible to AI.

4B computer-use model. LoRA adapters on a frozen Qwen3.5-4B, SFT then RL where the only reward is finishing the task. 92.9% vs 60.1% on GUI-360, no accessibility tree. Small enough to run on-prem for real SMB workflows. Pulling it onto the Spark this week