The "arms race" in the field of domestic large models is intensifying again. According to the latest news, the next-generation flagship model of Moonshot, Kimi K3, is expected to be officially launched in the third quarter of this year. As a rising star that has attracted widespread attention, every iteration of the Kimi series has captured the industry's attention. The launch of K3 is undoubtedly set to further raise the competition threshold for domestic large models.
2.5 Trillion Parameters Aimed at Industry Ceiling
In core technical indicators, Kimi K3 has demonstrated remarkable leapfrog growth. Its parameter scale is expected to reach 2.5 trillion, a figure that ranks among the top in the current domestic model landscape. In comparison, the parameter count of DeepSeek V4 Pro released recently is about 1.6 trillion, while Baidu's ERNIE 5.0 has around 2.4 trillion parameters. With a scale of 2.5 trillion parameters, Kimi K3 aims to comprehensively surpass its competitors in computing power and model capacity.
Evolution of a Million-Word Context Processing Capability
Aside from its massive parameter scale, Kimi K3 has also deepened its optimization in its signature "long text" advantage. The standard context length of the new model will be increased to approximately one million words, meaning users will gain a more coherent and in-depth interactive experience when handling extremely long documents, complex code repositories, or large-scale datasets.
Steady Progress: Moonshot's Business and Technology Landscape
Looking back at Moonshot's development journey, its iteration pace has been very tight. From achieving sub-second hot updates for a trillion-parameter model in September last year, to releasing the open-source thinking model Kimi K2 Thinking in November, and then K2.5 model topping the open-source list and achieving $100 million in annual recurring revenue (ARR) at the beginning of this year, the team has demonstrated strong engineering implementation capabilities.
Just recently, on April 21, Kimi released and open-sourced the K2.6 version, focusing on improving cluster processing capabilities. Only a week later, news about K3 followed closely. Moonshot's president, Zhang Yuting, once stated that the team is highly efficient in resource utilization. As K3 approaches its release, the competition for trillions of computing power has entered a feverish stage.
