PublishersMoonshot AI

Moonshot AI

3 documents read from this publisher.

moonshotai/Kimi-K3 · Hugging Face

first party release · model card · 2026-07-27 · 6 techniques

Configurable reasoning effort Normalized LatentMoE Preserved thinking history mode Context compaction MXFP4 Weights with MXFP8 Activations Kimi Delta Attention and Attention Residuals architecture

Kimi K3 Tech Blog: Open Frontier Intelligence

first party release · official blog · 2026-07-16 · 13 techniques

Quantization-Aware Training Sparse expert activation Attention Residuals (AttnRes)Normalized LatentMoE Preserved thinking history mode Kimi Delta Attention Per-Head Muon Quantile Balancing Context compaction Maximum thinking effort Fully Balanced Expert-Parallel Training KDA with prefill cache Vision in the loop

MoonshotAI/Kimi-K3

first party release · code repo · 14 techniques

Mixture of Experts Configurable reasoning effort Quantization-Aware Training Attention Residuals (AttnRes)Normalized LatentMoE Preserved thinking history mode Kimi Delta Attention Context compaction Gated Multi-Head Latent Attention MXFP4 Weights with MXFP8 Activations SiTU-GLU Always-on thinking mode MoonViT-V2 tool augmentation during evaluation