SourcesHugging Face community blog
publisher · 3 sources
Hugging Face community blog
3 sources read from this publisher, newest first, with the methods each one describes.
DeepSeek V4 Flash Is Now Official: What Changed in the 0731 Build
news analysis · third party analysis · 2026-07-31 · 5 methods
Kimi K3 Model Overview: 2.8T Parameters, MXFP4 Quantization, and What the Open Weights Mean for the Community
first party release · third party analysis · 2026-07-17 · 15 methods
MiniMax Goes Sparse: Decoding M3's Attention from a Single Diagram
news analysis · third party analysis · 2026-05-29 · 16 methods
Mixture of expertsDeepSeek Sparse AttentionGrouped-query attentionSparse attentionCompressed Sparse AttentionMulti-head Latent AttentionTop-k block selectionBlock Max PoolingBlock-level selectionBlock-sparse grouped-query attentionDense Attention FallbackLightning AttentionNative Sparse AttentionNatively trained sparsityRoPE with attention sinkSingle-head index key