I wanted to see how far I could push the "Flash" in Gemini 3.1. Most AI interfaces feel bloated and slow, so I built Nano Banana 2 to focus on raw speed.

Technical Highlights:

Stack: Next.js + Edge Runtime to minimize TTFB.

Optimization: Custom streaming pipeline for Nano Banana 2 image preview tokens.

Locales: Automated i18n for 13 regions (MS, IL, PL, etc.) based on real-time search breakout data.

The goal was to build a globally responsive UI that feels native regardless of where the user is. Would love your feedback on the latency and the streaming implementation.

Free to try for the HN community: https://nano-banana2.me/