Moonshot AI has unveiled Kimi K2, a groundbreaking open-source language model poised to redefine the landscape of AI-driven tool usage and agentic tasks. With a staggering 1 trillion parameters and a 32 billion active mixture-of-experts (MoE) architecture, Kimi K2 exhibits superior performance that rivals the previous impact of DeepSeek R1. The open-weight design not only boosts reliability in function calling compared to existing models like Claude and GPT but also opens avenues for generating synthetic training data to enhance tool-calling capabilities in other models.
Offering two distinct variants—K2-Base for fine-tuning applications and K2-Instruct for general chat—the model is trained using the innovative Muon optimizer across 15.5 trillion tokens. This extensive training allows Kimi K2 to achieve exceptional results in coding, mathematics, and reasoning benchmarks, often setting new standards among open-source models. With support for a 128K context length and deployment via popular inference engines like vLLM and SGLang, Kimi K2 stands out with its advanced agentic capabilities for autonomous tool usage and multi-step task completion.
Kimi K2’s performance on coding benchmarks is noteworthy, surpassing GPT-4 in key areas such as LiveCodeBench (53.7% vs. 44.7%), SWE Bench Verified, Tau2, and AceBench. Despite its slower speed, the model’s robust functionality and competitive API pricing ranging from $0.15 to $2.50 per million tokens make it an attractive option for developers and businesses alike, challenging the commercial strategies of firms like OpenAI and Anthropic.
While the model currently lacks multimodal and thought-mode capabilities, Kimi K2’s state-of-the-art results in agentic tasks mark a significant leap forward in AI development. Despite some licensing restrictions for large commercial users, its open-source nature promises to accelerate the evolution of more capable agentic AI systems across the industry.

Collections

Moonshot AI released Kimi K2, an open-source language model with 1 trillion parameters and 32 billion active mixture-of-experts architecture. The model offers two variants (K2-Base and K2-Instruct) and excels in coding, mathematics, and reasoning benchmarks, outperforming GPT-4 in several areas including LiveCodeBench. With 128K context length support and competitive API pricing ($0.15-$2.50 per million tokens), Kimi K2 demonstrates superior agentic capabilities for autonomous tool usage and multi-step task completion, though it currently lacks multimodal features.

Kimi K2: Open-Source AI Model Redefining Coding Benchmarks