Kimi K2 is a groundbreaking mixture-of-experts language model developed by MoonshotAI, featuring 32 billion activated parameters and 1 trillion total parameters. Designed for exceptional performance in frontier knowledge, reasoning, and coding tasks, Kimi K2 is optimized for agentic capabilities, enabling it to act and solve problems autonomously.
Kimi K2 is a state-of-the-art AI model that leverages the mixture-of-experts architecture to achieve unprecedented scale and performance. With 1 trillion total parameters and 32 billion activated parameters, it represents a significant advancement in AI technology.
Utilizes a mixture-of-experts (MoE) approach with 384 experts, allowing for efficient and scalable training.
Pre-trained on 15.6 trillion tokens with zero training instability, ensuring robust and reliable performance.
Employs the MuonClip Optimizer to enhance token efficiency and prevent gradient explosions.