Google quietly releases COSMO experimental AI assistant app with local Gemini Nano

Google published an experimental AI assistant application called COSMO on the Play Store on April 30, 2026. The 1.13 GB app, which appears to originate from Google Research (package name: com.google.research.air.cosmo), includes a local Gemini Nano model that runs entirely on-device.

14 automated skills

COSMO includes 14 "Skills" or automated capabilities:

• List Tracker: Automatically suggests Google Keep lists
• Document Writer: Offers to create documents, letters, or summaries
• Calendar Event Suggester: Schedules agreed-upon plans automatically
• Browser Agent: Automates tasks using "Mariner"
• Add Timer: Creates timers in the clock app for time-bound tasks
• Deep Research: Generates full reports from multiple sources
• Quick Photo Lookup: Finds photos without interrupting conversations
• Google it: Performs quick web searches for questions
• Jargon Definitions: Explains complex terminology and acronyms
• Provide Insight: Offers ideas and information
• People Understanding: Provides context about people
• Event Understanding: Provides context about events
• Recall: Shows things users are trying to remember
• Conversation Summary: Summarizes recently ended conversations

Hybrid on-device and cloud processing

The app includes three "Fulfillment Model" options in settings:

1. Hybrid: Uses "PI" when online, Nano offline
2. PI Only: Server-based processing only
3. Nano Only: Local on-device processing only

The meaning of "PI" remains unclear, though the local option explicitly refers to Gemini Nano. Additional features include voice match setup and screen access permissions.

Likely premature release

The app presents a basic chat interface that reflects its experimental nature. According to 9to5Google, the release appears premature or accidental ahead of Google I/O 2026 later in May. The app was published on Google's main Play Store account rather than a separate experimental channel.

What this means

COSMO represents Google's exploration of proactive, on-device AI assistance that extends beyond simple chat interfaces. The inclusion of a 1.13 GB local model signals Google's commitment to privacy-focused on-device processing, while the hybrid approach allows fallback to cloud capabilities when needed. The reference to "Mariner" in the Browser Agent skill suggests integration with Google's previously announced web automation capabilities. This is clearly a research preview rather than a consumer product, likely testing infrastructure and interaction patterns for future Gemini integration across Android.