GUI agents that tap through apps and phone home when stuck

What it does MAI-UI is a suite of foundation models—2B, 8B, 32B, and 235B-A22B parameters—that operate GUIs by looking at the screen and emitting taps, swipes, and text. It handles real-world tasks like buying groceries, booking trains, and cross-app shopping (searching Xiaohongshu and buying on Taobao). The open-source release includes 2B and 8B weights on Hugging Face, plus two Jupyter notebooks that wrap the model into MAIGroundingAgent and MAIUINavigationAgent.

The interesting bit The device-cloud collaboration system is the clever part: the agent evaluates the task execution state and data sensitivity, then either runs locally or hands off to a 235B cloud model. It also speaks MCP, letting it invoke external tools like AMap mid-workflow instead of fumbling through maps manually.

Key highlights

• Sweeps ScreenSpot-Pro across all size categories (2B, 8B, 32B), with the 32B scoring 67.9% without zoom-in tricks.
• Tops AndroidWorld at 76.7% and MobileWorld at 41.7% for end-to-end vision agents.
• Trained with online RL scaling up to 512 parallel environments and 50-step context windows.
• Ships inference-ready 2B and 8B weights; the 32B and 235B record-holders are not yet publicly released.

Caveats

• The repository is largely quickstart glue—two cookbook notebooks and demo GIFs—rather than the full training framework or agent runtime.
• Only 2B and 8B weights are downloadable; the 32B and 235B models that set several records are not publicly released.
• Requires pinned vLLM 0.11.0 and trust-remote-code, suggesting a non-standard model architecture.

Verdict Worth a look if you’re building mobile automation or benchmarking GUI agents; skip it if you need training code or a fully open large model.