The tech world is abuzz with the recent launch of the SiliconCloud platform's Kimi K2, an open-source model developed by the Moon's Dark Side. This colossal model boasts a staggering 1 trillion parameters and 32 billion active parameters, positioning it as a formidable contender against top-tier commercial models. Kimi K2 excels in encoding and general agent tasks, offering developers an incredibly powerful tool at their disposal.
With the ability to support up to 128K maximum context length, Kimi K2 is not only a performance powerhouse but also competitively priced. The input cost is set at 4 yuan per million tokens, and the output cost is 16 yuan per million tokens. SiliconCloud has further expanded its offerings by introducing the Kimi-Dev-72B programming model.
Users across the globe can experience the prowess of Kimi K2 through SiliconCloud's domestic and international sites. Upon registration, new users are greeted with a 14-yuan trial credit, facilitating their first foray into the platform's capabilities. Domestic users can access the model via the [domestic site link](https://cloud.siliconflow.cn/models), while international users can explore it through the [international site link](https://cloud.siliconcom/models).
Kimi K2's technical prowess is rooted in three main advantages: First, its large-scale training on 15.5 trillion tokens ensures model stability. Second, the MuonClip optimizer allows for stable scaling in extensive environments. Lastly, Kimi K2 is an agent-driven model, specifically designed for tool invocation, logical reasoning, and autonomous problem-solving.
Benchmark tests reveal Kimi K2's exceptional capabilities in autonomous programming, tool invocation, and mathematical reasoning. The Moon's Dark Side asserts that Kimi K2 lays a solid foundation for building general agent capabilities, with future updates promising the addition of thinking and visual understanding abilities.
The response to Kimi K2 has been overwhelmingly positive. Aravind Srinivas, founder of Perplexity, lauds its outstanding performance in internal evaluations and anticipates subsequent training based on this model. Users have also reported Kimi K2's remarkable ability to complete web-based Minecraft tasks in one go, outperforming other models significantly.
Beyond its coding prowess, Kimi K2 has garnered praise for its linguistic abilities, demonstrating a more diverse vocabulary. Users are satisfied with Kimi K2's overall performance and eagerly await future updates and expansions.
Embrace the future of AI with Kimi K2, a model that sets new standards in performance and versatility.