PublishersHugging Face community blog

Hugging Face community blog

3 documents read from this publisher.

DeepSeek V4 Flash Is Now Official: What Changed in the 0731 Build

news analysis · third party analysis · 2026-07-31 · 5 techniques

Configurable reasoning effort DSpark A/B testing on real tasks cache-hit cost modeling for agent workloads Re-post-training for agentic capabilities

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 techniques

Mixture of Experts Quantization-Aware Training Attention Residuals (AttnRes)Normalized LatentMoE Per-Head Muon Quantile Balancing Gated Multi-Head Latent Attention MXFP4 weight quantization KDA MXFP8 activation quantization Hybrid linear attention Latent-space routing Mooncake SiTU Soft dropping

MiniMax Goes Sparse: Decoding M3's Attention from a Single Diagram

news analysis · third party analysis · 2026-05-29 · 16 techniques

Mixture of Experts DeepSeek Sparse Attention Grouped-query attention Compressed Sparse Attention Sparse attention Multi-Head Latent Attention Top-k block selection Block Max Pooling Block-level selection Block-sparse grouped-query attention Dense Attention Fallback Lightning Attention Native Sparse Attention Natively trained sparsity RoPE with attention sink Single-head index key