Faiss 1.16.0 adds e-means for large-scale clustering

What is changing?

The `faiss-cpu 1.16.0` package was uploaded to PyPI through Trusted Publishing on November 10. Although the GitHub Releases page and main changelog have not yet caught up, the package record, e-means code in the official Faiss repository, and the November 6 research paper confirm the release and the feature’s scope.

Instead of rebuilding sums and counts over the full dataset on every pass as Lloyd k-means does, e-means updates exponentially weighted sums and counts from mini-batches. It can revive underused centroids by splitting active ones and finalize centroids from assignments collected during the last pass. Faiss 1.16.0 includes NumPy and PyTorch paths, CPU/GPU assignment, multi-GPU training, and centroid sharding.

The researchers report that e-means exceeded the loss reached by Lloyd’s method after 25 epochs in 1.4 to 2.6 epochs on the compared datasets, with 10–45× GPU wall-clock speedups in their comparisons. These are provider research results; hardware, dimensionality, cluster count, and approximate-assignment choices determine real gains.