Gemma 4 hits 81 tok/sec on M5 Max

Google's A4B architecture, activating 4 billion of its 26 billion total parameters, allows the M5 Max's 614 GB/s bandwidth to deliver inference speeds formerly reserved for 7B-class models. This 81 tokens per second performance provides the ultra-low latency required for complex, multi-step agentic tool-calling without frustrating wait times. While the 114W peak power draw is impressively efficient, thermal throttling remains a consideration for extended generation sessions. Apple's unified memory architecture continues to be a major advantage, allowing 26B weights to be loaded without the VRAM bottlenecks typical of consumer Nvidia mobile GPUs.