100+ agents collaborated for a week to make Gemma 4 run 5x faster in vLLM

Thomas Wolf calls multi-agent collaboration one of the most interesting agent behaviors right now — and backs it with an experiment in which 100+ agents, left to openly collaborate for a week, wrung a 5x inference speedup out of Gemma 4 in vLLM. Maziyar Panahi quote-tweets to aim that same energy at on-device clinical AI (OpenMed's 1,500+ locally-running models), where you want the speed but can't trade away accuracy.

@MaziyarPanahi on X →