Google launches the AI Edge Gallery app, AI runs completely offline on mobile phones.

Google has quietly issued AI Edge Gallery since May 31, an AI application placed directly on smartphones, without the need for internet or sharing data with Big Tech servers.

• Open source app (Apache 2.0) is available on GitHub, starting with Android, iOS coming soon.
• Run AI models like Gemma 3n completely offline, processing from image analysis to coding using only phone hardware.
• Main features include: AI Chat (conversation), Ask Image (image analysis), Prompt Lab (single tasks such as rewriting text).
• Users can download models from Hugging Face, but are limited to the .task format, which does not support as much as competitors.
• Performance: The small model runs fast and smoothly (~20 tokens/second CPU, over 100 tokens/second GPU), meets expectations well like the first generation GPT-3.5.
• Major advantage: no network needed, high security, suitable for users who need to keep data private such as doctors, journalists.
• Limitations: large memory ( large model approximately 4.4 GB), context window limit 4096 tokens, model diversity is low, GPU sometimes less accurate than CPU.
• No need to worry about strict censorship like central services, can use a simple jailbreak.
• Save costs as there are no cloud fees, only using phone resources.

In summary, the app is suitable for those who need AI security, do not connect to the network, or want to test AI models offline, although the speed and experience cannot yet compare to modern cloud AI services.