Xiaomi opens MiMo Desktop beta for editable work and computer control

Xiaomi MiMo Desktop has opened for invitation testing, bringing document creation, live previews and computer interaction into a desktop agent application. Xiaomi’s MiMo site describes early access for office work, design, coding and multimodal creation, with limited-time access to next-generation models.

The invitation page currently excludes the UK, EU countries and South Korea, alongside certain other countries. That regional restriction is the first practical check for anyone considering the beta.

A deliverable that can be revised

Xiaomi describes a workflow that takes files as context and produces items such as slides, web pages, applications and 3D work. Results can be operated in the session or preview pane. Users can select a local area for revision and return to an earlier version.

For a reader preparing a presentation from a folder of documents, the meaningful outcome is an editable file with usable structure. A live preview adds a second check: whether the deliverable behaves as intended when someone operates it. These are useful acceptance criteria for testing the beta with a real task.

Browsing, operating and checking

The advertised browser controls cover research, extraction and form interaction. The global version also describes screen reading, keyboard and mouse control, plus recording a workflow for reuse. Xiaomi says the system checks results after key steps and can revise its actions or stop.

Xiaomi also advertises dynamic model selection, targeted regeneration and caching as ways to manage long tasks. Cache-hit rates describe reused computation; they should be read separately from a user’s final bill.

A sensible first task is bounded and easy to inspect: a short presentation, a spreadsheet summary or a small website using copies of source files. Check the output, operate its controls, and inspect what changed during a revision. The beta’s central proposition is a complete cycle from supplied material to editable work that a user can test.