Player not loading? Watch on YouTube
In this official Google product-team conversation, Google AI Edge Gallery is presented as an app for trying Gemma 4 on a phone. Alice Zheng explains that inference runs on the device, and the app is free on Android and iOS. Its code is open source on GitHub. The discussion introduces 2B and 4B models aimed at phones, alongside 26B and 31B models aimed at laptops, desktops and servers.
The examples show how to use a local LLM for everyday tasks. An offline chat prompt asks for an explanation of why the sky is blue for a five-year-old, without search grounding. An audio demo turns a spoken shopping memo into a structured list. Other suggested uses include translating photographed menus and extracting book titles and authors into JSON.
Agent skills add tools and task instructions. The speakers describe Wikipedia and WebFetch access for information outside the model's training knowledge, plus HTML and JavaScript output for interactive content. These web examples involve external information access, even though the model runs on the phone. Users can load skills from a URL or local file and inspect community examples.
The preview covers experimental MCP integration planned for Android first, then iOS, and persistent chat history. Olivier Lacombe says this year's 2B model matches last year's 27B dense model in performance, but the discussion does not specify a benchmark or device requirements for that comparison.
