Gemma 4 hits 3x faster local inference via speculative decoding; 'AI Ultra Lite' tier and Gemini Agent prep for I/O

Google's open Gemma 4 models use multi-token-prediction speculative decoding for up to 3x faster local inference, with the largest variant runnable on a single high-power accelerator. Google is also readying a 'Gemini Agent' upgrade as a 24/7 digital partner, plus a cheaper 'AI Ultra Lite' plan with explicit usage caps for Gemini, all positioned ahead of Google I/O.