Moonshot AI's Kimi K3 is a native multimodal Mixture-of-Experts (MoE) architecture featuring 2.8 trillion total parameters, with 104 billion activated parameters per token. To facilitate a 1-million-token context window, the model utilizes a 3:1 hybrid attention mechanism consisting of three Kimi Delta Attention (KDA) layers followed by one Gated Multi-Head Latent Attention (MLA) layer per block. KDA employs a channel-wise forget gate with a lower-bounded log-decay parameterization ($g_{min} = -5$) to ensure numerical stability within the BF16 dynamic range during chunkwise parallel execution. To mitigate signal dilution in its 93 layers, Kimi K3 implements Attention Residuals (AttnRes), allowing layers to attend to preceding depth representations via pseudo-queries, partitioned into eight 12-layer blocks to reduce memory overhead from $O(Ld)$ to $O(Nd)$. The model scales width through 896 routed experts (16 activated per token) and two shared experts, utilizing Normalized LatentMoE and SiTU-GLU to stabilize training. Routing is optimized via Quantile Balancing (QB), which adjusts router biases using global batch quantiles. For efficiency, the model uses MXFP4 weight quantization and a specialized serving engine that packs KDA states and MLA KV blocks into paged memory, achieving a 2.5x efficiency improvement over Kimi K2.