Experiments with On-device AI — What building on Gemini Nano actually teaches you

Short summary

A developer shares hard-won lessons from building a Chrome extension using Gemini Nano's on-device AI APIs. Key challenges include managing feature-flag-gated task APIs with fallback chains, handling multi-state availability (unavailable/downloadable/downloading/available), and accounting for a much smaller model than cloud APIs. Practical code examples show how to build resilient on-device AI features that work across real user machines.

• •On-device AI APIs are flag-gated and roll out independently — build fallback chains per capability
• •availability() returns four states, not boolean; first calls trigger slow model downloads
• •Small on-device models make length/count instructions approximate — don't display raw output

Generated with AI, which can make mistakes.