Meta FAIR releases e-means, a k-means replacement in FAISS 1.16.0

Meta FAIR's e-means is a mini-batch k-means replacement included in FAISS 1.16.0. It is described as reaching Lloyd's 25-epoch loss in 1.4 to 2.6 epochs.

The method is reported to give 10 to 45 times GPU speedups and needs no extra hyperparameters. The figures come from the announcement and have not been independently checked here.

If your agent pipeline already clusters embeddings with FAISS, e-means could cut the number of passes needed to reach a comparable clustering loss. Fewer epochs and no extra tuning knobs would make it simpler to swap in for mini-batch k-means.

Benchmark gains on GPU depend on the data and hardware, so run your own comparison before changing a production job.