Moonshot AI, backed by Alibaba, has launched Kimi K2, a new open-source AI model designed for advanced agentic tasks. The model boasts 1 trillion parameters, with 32 billion activated during use, and excels in areas like coding, maths, and complex problem-solving. Kimi K2 comes in two versions: Kimi-K2-Base for researchers and Kimi-K2-Instruct for general use. It is freely available via web and mobile applications.
Kimi K2 employs a Mixture-of-Experts (MoE) architecture, dividing the model into specialised sub-networks to improve efficiency and reduce computational costs. This approach enables the model to understand how to utilise tools and complete practical tasks, such as generating salary analyses with statistical visualisation. The release reflects a growing trend towards open-source development in the AI sector, allowing for wider adoption and collaborative improvements.
Interestingly, shortly after Kimi K2's release, OpenAI CEO Sam Altman announced a delay in their own planned open-source AI model, citing the need for further safety testing. Moonshot AI's Kimi K2 is expected to draw significant attention in the global AI market, especially after DeepSeek's open-source V3 and R1 models.
