E-means is a k-means variant from Meta FAIR, presented as a replacement for Lloyd's algorithm. It updates centroids with an exponential moving average over mini-batches rather than full passes over the data, sets batch size and step size automatically from scaling laws fitted on the problem's size, revives dead clusters by splitting large ones, anneals the step size on a cosine schedule, and ends with a finalization stage that lands on each cluster's exact last-epoch mean. The authors report reaching Lloyd's 25-epoch test loss in 1.2 to 3 epochs (10-45x faster on GPU, 8-15x on CPU) while also reaching a lower loss, and escaping local minima much more often. Like Lloyd's, it targets static datasets, unlike the constantly moving distributions of VQ-VAE codebook learning. It ships with Faiss.