Kimi K2

Kimi K2 pushed open-weight scale to a trillion parameters: Moonshot AI's MoE series activates 32B parameters per token across 384 experts with 8 selected plus one shared expert, 61 layers, MLA attention, a 160k vocabulary, and a 128K context window. Training stability at that scale came from applying the MuonClip variant of the Muon optimizer. K2-Base serves fine-tuners while K2-Instruct is post-trained for chat and agentic tool use, and the later K2 Thinking release added an open reasoning model that Moonshot reports sustaining 200 to 300 sequential tool calls without human intervention. Weights carry a Modified MIT license: fully commercial, with the sole condition that products exceeding 100 million monthly active users or 20 million US dollars in monthly revenue must prominently display Kimi K2 in their interface. Self-hosting the 1T checkpoint requires a multi-GPU cluster served through vLLM, SGLang, KTransformers, or TensorRT-LLM, so many teams instead use the OpenAI- and Anthropic-compatible API on Moonshot's hosted Kimi platform. It set state-of-the-art open results on agentic and coding benchmarks at release.