Agents Collaborate to Make Gemma 4 Inference 5x Faster, Peaking at 491.8 TPS

In the "Fast Gemma Challenge" run by Hugging Face and Google, more than 100 AI agents and humans collaborated over six days to roughly quintuple the inference speed of google/gemma-4-E4B-it on a single NVIDIA A10G GPU. The overall fastest result reached 491.8 tokens per second (TPS), though it came with a drop in model quality in other areas. The fastest lossless result that preserved quality was 315 TPS.