Moonshot AI unveils Kimi K3 model with 2.8 trillion parameters

Moonshot AI, the developer of the Kimi chatbot, announced its largest model yet, Kimi K3, with 2.8 trillion total parameters. The model employs a Mixture of Experts (MoE) architecture, activating only a subset of parameters per query to balance performance and efficiency. Moonshot claims Kimi K3 achieves performance on par with GPT-4o and DeepSeek-V3 in benchmarks such as MMLU and HumanEval. The company emphasized that the model's deployment cost is significantly lower than competing models of similar capability, due to the MoE design. Kimi K3 is designed to power the Kimi assistant and will be available via API, with pricing details to be announced. Moonshot also noted that the model was trained on a cluster of 10,000 GPUs and required substantial investment, but the company's accounting suggests that the long-term cost of serving the model will be offset by its efficiency.