The release of Kimi K2 has sparked widespread discussion, surpassing many leading large language models.

Kimi K2, a MoE-based foundation model released by Chinese AI unicorn Moonshot AI, has garnered widespread attention within just two days of its launch. On the API aggregation platform OpenRouter, its token usage quickly surpassed that of xAI’s Grok 4. With a total parameter count of 1 trillion and 32B active parameters, Kimi K2 demonstrates outstanding performance even on non-NVIDIA hardware, while operating at only 20% of Claude 4’s cost. Developers have praised its versatile coding capabilities and robust API services, noting its strong real-world performance. Additionally, Kimi K2 shares architectural similarities with DeepSeek and has made notable advancements in optimization and agentic capabilities, being hailed as another major victory for open-source models.