Kimi K3 Open-Weight Model

Overview

Kimi K3 is a frontier Chinese open-weight model with about 2.8 trillion parameters and a Mixture‑of‑Experts design built on Kimi Delta Attention, Attention Residuals, and innovations like MoonViT‑V2, Hybrid KDA‑MLA, Stable LatentMoE, AttnRes, and Per‑Head Muon. It supports a 1M-token context window for extremely long conversations and documents, and handles both text and images for long‑form coding, debugging, research, planning, and agent‑style workflows. The model is being adopted across clouds and enterprise AI platforms, making it attractive for teams that want strong Chinese multimodal performance with open weights.