Google DeepMind’s Gemma team and Hugging Face ran the “Fast Gemma Challenge,” achieving a 5x improvement in inference speed for the Gemma 4 model through more than 100 AI agents and human participants working together over roughly six days. The challenge ran from June 26 to July 2, targeting the Gemma 4 E4B-it model on a single NVIDIA A10G GPU with 24 GB of memory. Results announced on July 10 showed an absolute peak performance of 491.8 tokens per second and a best “lossless” submission at 315 TPS, both representing a roughly 5x jump from baseline. Techniques included the vLLM inference engine, speculative decoding, quantization, IPC/ZMQ tuning, and integration with TGI. Making inference 5x faster on the same hardware means serving 5x more users at the same cost or cutting compute bills by 80% for the same workload. The single-GPU constraint was deliberate, as A10G cards are widely available through cloud providers like AWS. The challenge crowdsourced optimization with AI agents as first-class contributors alongside human engineers.