iPhone 17 Pro tops Gemma 4 benchmarks

Apple's silicon lead is now a bottleneck for Google's own models on Android hardware. The A19 Pro's unified memory and neural engine efficiency allow Gemma 4 E2B to run at near-instant speeds of 40 tokens per second, while the Pixel 10 Pro struggles with the same workload. Native multimodality and agentic capabilities in Gemma 4 make this speed difference critical for real-time speech and vision applications. Android OEMs are falling behind in the "token war," which may push developers to prioritize iOS for high-performance on-device AI features despite Gemma being a Google product. Gemma 4's Apache 2.0 license and 128K context window make it a prime candidate for local RAG, where the iPhone's 12GB RAM provides significant headroom for complex document analysis.