Meta swaps k-means for a faster clustering step in FAISS 1.16

single source· 1 articles · confidence: low · first seen 2026-11-10 23:20 UTC

What this means for you

If your pipeline calls FAISS's mini-batch k-means, the replacement ships in 1.16.0 and, according to Meta, adds no hyperparameters to tune. The 10-45x speedup is Meta's own figure; rerun your own clustering before banking on it.

Meta FAIR has replaced the mini-batch k-means routine in FAISS with a variant it calls e-means, shipped in version 1.16.0. k-means groups data by repeatedly assigning points to the nearest of k centres and recomputing those centres; an epoch is one pass over the data. Meta reports that e-means reaches the loss the standard Lloyd's algorithm reaches after 25 epochs in 1.4 to 2.6 epochs, with 10-45x GPU speedups and no new hyperparameters to tune. The figures are Meta's own and no evaluation date is given.

Key facts

  • ·e-means ships in FAISS 1.16.0 as a replacement for mini-batch k-means source
  • ·Meta reports e-means reaches Lloyd's 25-epoch loss in 1.4 to 2.6 epochs source
  • ·Meta reports 10-45x GPU speedups over mini-batch k-means source
  • ·The method requires no additional hyperparameters, per Meta source
  • ·No evaluation date is given for the reported epoch counts or speedups source

What the sources say

  • MarkTechPost — Brief write-up of the FAISS release, giving the epoch and speedup figures without an evaluation date.

Sources

The original reporting. Follow these — they did the work.

← the wire