2.8T total, 104B active (MoE)
1,048,576 tokens
Kimi K3 License
Moonshot AI's open-weight native multimodal agentic flagship for long-horizon coding, reasoning, knowledge work, visual tasks, and tool-driven workflows. Kimi K3 is a 93-layer, 2.8-trillion-parameter Mixture-of-Experts model with 104B activated parameters, making it the first publicly released open-weight model in the 3T parameter class. Its architecture combines 69 Kimi Delta Attention layers with 24 gated Multi-head Latent Attention layers, Attention Residuals for information flow across model depth, and Stable LatentMoE. The MoE contains 896 routed experts with 16 selected per token plus two shared experts, alongside one dense layer, a 7,168-dimensional attention hidden state, 96 attention heads, a 160K vocabulary, and the SiTU-GLU activation function. Moonshot says these changes provide approximately 2.5 times the overall scaling efficiency of Kimi K2. Native multimodal support covers text, images, and video through a 401M-parameter MoonViT-V2 vision encoder, while the 1,048,576-token context window targets enormous repositories, document collections, extended agent trajectories, and multimodal knowledge work. The released checkpoint was quantization-aware trained from the supervised-fine-tuning stage onward with MXFP4 weights and MXFP8 activations and occupies approximately 1.56 TB on Hugging Face. Post-training includes reinforcement learning across general, coding, agentic, and multiple reasoning-effort settings, supported by infrastructure for million-token agentic RL with persistent rollout and sandbox state. Kimi K3 always operates in thinking mode, exposes low, high, and max reasoning-effort levels with max as the default, and requires preserved reasoning history across multi-turn conversations and tool calls. Moonshot's maximum-effort evaluations report 93.5 on GPQA Diamond, 67.5 on DeepSWE, 88.3 on Terminal-Bench 2.1, 77.8 on ProgramBench, 81.2 on FrontierSWE, 42.0 on SWE-Marathon, 91.2 on BrowseComp with context management, 84.2 on MCP-Atlas, 63.3 on OfficeQA Pro, 91.1 on OmniDocBench, and 90.0 on Video-MME with subtitles. Official deployment guidance covers vLLM, SGLang, and TokenSpeed. The custom Kimi K3 License broadly permits use, modification, distribution, fine-tuning, and sale, but commercial model-as-a-service operators exceeding $20 million in aggregate revenue over a consecutive 12-month period must obtain a separate Moonshot agreement; very large commercial products also carry a Kimi K3 display requirement.