MoonshotAI: Kimi K2 0905

Kimi K2 0905 is the September update of Moonshot AI's Kimi K2 series, a large-scale Mixture-of-Experts (MoE) language model. Boasting 1 trillion total parameters with 32 billion active per forward pass, it significantly extends its long-context inference capability up to 256k tokens. This model is engineered for superior agentic capabilities, including advanced tool use, sophisticated reasoning, and efficient code synthesis. This update brings notable improvements to agentic coding, offering higher accuracy and better generalization across various scaffolds. It also enhances frontend coding, producing more aesthetic and functional outputs for web, 3D, and related tasks. Kimi K2 0905 demonstrates exceptional performance across critical benchmarks like LiveCodeBench, SWE-bench for coding; ZebraLogic, GPQA for reasoning; and Tau2, AceBench for tool-use. The model is trained with a novel stack incorporating the MuonClip optimizer for stable large-scale MoE training. With a generous context window of 262K tokens and a max output of 4K tokens, Kimi K2 0905 is ideal for complex, long-form tasks. It supports functions, code generation, and streaming. Access this PRO tier model on Multi AI for $0.39/1.90 per 1M tokens (input/output).