GoKawiilMeta FAIR open-sources e-means, a faster alternative to Lloyd's k-means algorithm
Meta FAIR has released e-means, a new k-means clustering algorithm that uses mini-batch updates with exponential moving averages instead of the decades-old MacQueen or Lloyd update rules. The company reports it reaches lower test loss than Lloyd's algorithm in 1.2 to 3 epochs versus 25, translating to 10–45x faster wall-clock performance on GPUs and 8–15x faster on CPUs compared to existing implementations like Flash-KMeans and Faiss.
GoKawiil's interpretation of the reporting above, not reported fact.
K-means clustering underlies many data processing and machine learning pipelines, so a method that improves both speed and accuracy simultaneously—rather than trading one for the other—could reduce compute costs at scale. Meta's choice to open-source the algorithm suggests it wants the broader research and engineering community to adopt and validate it, potentially influencing how clustering tools are built going forward.
- E-means reportedly cuts epochs needed by 8-20x compared to Lloyd's algorithm while achieving lower loss
- Wall-clock speedups range from 10-45x on GPU and 8-15x on CPU versus existing tools
- The algorithm applies modern mini-batch optimization techniques to a classic, often full-batch-optimized problem
Source: e-means.github.io — Pierre-Emmanuel Mazaré, 2026-11-09
Published there as: “E-means: 10x faster k-means at equal loss”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.