Gemma 4 26B tops 600 tok/s

This is a strong reminder that speculative decoding can turn a fast open model into something that feels genuinely server-grade on consumer hardware, but the result is still a narrow single-request benchmark, not a universal production guarantee.

• –The headline gain is real: 2.56x throughput on one GPU is enough to change how practical 26B-class serving feels.
• –The serving lesson matters more than the raw peak: `4096` had slightly better mean latency, but `8192` cleaned up the tail, which is what users actually feel.
• –The benchmark used a random dataset with `concurrency=1` and `request rate=1`, so real chat, code, or tool-use traffic may accept speculative tokens very differently.
• –If DFlash holds up on realistic workloads, it becomes a credible way to stretch RTX 5090-class boxes into serious local inference servers.