E-means combines minibatch clustering with exponential moving averages and automatic parameter selection to reduce k-means iteration costs. The authors report lower loss than 25 Lloyd iterations after roughly one to three data passes, with GPU speedups of 10–45 times and CPU speedups of 8–15 times under their test conditions. The paper also describes clustering 10 billion DINO vectors into 16 million groups. A Faiss implementation is available, but these author-reported results require validation of both quality and runtime on production workloads.