tl;dr: Moonshot AI released Kimi K3, an open-weight 2.8T-parameter MoE model (104B activated) with native multimodality and a 1M-token context window, built on new Kimi Delta Attention (KDA) and Attention Residuals architecture with MXFP4 quantization-aware training. Benchmarks show it competitive with Claude Opus 4.8, GPT-5.6, and Claude Fable 5 across reasoning, coding, agentic, and vision tasks, positioning it as the first open 3T-class frontier model. Weights are available on HuggingFace under the Kimi K3 License, with vLLM/SGLang deployment support and an OpenAI/Anthropic-compatible API.