Kimi K3
Moonshot AI · 2026-07-16
Kimi K3 is Moonshot AI's flagship reasoning model released on July 16, 2026, an approximately 2.8-trillion-parameter Stable LatentMoE transformer that activates 16 of 896 experts and supports a 1,048,576-token context with native image and video input. Its architecture combines Kimi Delta Attention, Attention Residuals, Gated MLA, and Sigmoid Tanh Unit, while Per-Head Muon and SFT-stage quantization-aware training with MXFP4 weights and MXFP8 activations support scaling; Moonshot reports roughly 2.5× better scaling efficiency than Kimi K2. K3 always reasons with low, high, or max effort (max at API launch), and Moonshot released the full model weights on July 27, 2026, under the Kimi K3 License.