Chinese startup Moonshot AI, backed by Alibaba, has launched Kimi K2, the world’s first open-source trillion-parameter LLM (with 32 billion active at inference). Based on a Mixture-of-Experts architecture and trained on 15.5 trillion tokens using the novel MuonClip optimizer, Kimi K2 achieves impressive stability and scaling. It surpasses GPT‑4.1 on programming benchmarks, scoring 65.8% on SWE-bench Verified and 53.7% on LiveCodeBench, close to Anthropic’s Claude Sonnet 4.
Designed for tool use and autonomous agents, it also posts a strong 76.5% on AceBench. Available under MIT license (with attribution for commercial use), it can run locally or via API compatible with OpenAI’s. Moonshot’s aggressive pricing — around $0.15 per million input tokens and $2.50 per million output — puts competitive pressure on closed models. This milestone could reshape the global AI race, positioning Chinese firms as serious challengers to OpenAI and Anthropic.
