Kimi K2 0905 is a large-scale Mixture-of-Experts (MoE) language model developed by Moonshot AI. Released in September 2025 as a refined iteration of the Kimi K2 series, it is designed with a specific focus on agentic intelligence, featuring enhanced capabilities for autonomous tool use, multi-step reasoning, and complex code synthesis. The model comprises approximately 1 trillion total parameters, though it maintains computational efficiency by activating only 32 billion parameters per token during inference.
A primary advancement in the 0905 update is its expanded context window, which supports up to 256,000 tokens. This capacity allows the model to process and reason over extensive documentation, large-scale codebases, and long-horizon tasks more effectively than earlier versions. Technical optimizations include the use of the MuonClip optimizer for training stability at scale and the Multi-head Latent Attention (MLA) mechanism for efficient processing.
Kimi K2 0905 has demonstrated high performance in software engineering benchmarks, particularly on the SWE-bench leaderboard. It is utilized both as a general-purpose assistant and as a foundation for specialized agentic workflows, such as frontend web development and automated debugging, where it provides high instruction fidelity and structured outputs.