Recently, the SiliconCloud platform officially launched the MoE model Kimi K2, an open-source model from Moonshot AI. This model has a total of 1T parameters and 32B activated parameters, performing well in various benchmark performance tests, comparable to top commercial models. Kimi K2 is particularly good at coding and general agent (Agent) tasks, providing developers with a powerful tool.
Kimi K2 supports a maximum context length of up to 128K, with competitive pricing: input cost is 4 yuan per million Tokens, and output cost is 16 yuan per million Tokens. In addition, SiliconCloud also launched another programming large model, Kimi-Dev-72B, further expanding its product line.
Users can experience Kimi K2 through SiliconCloud's domestic and international sites. New users will automatically receive 14 yuan of trial credit after registration, making it convenient for them to try the model on the platform. Domestic users can use the model via [Domestic Site Online Experience Link](https://cloud.siliconflow.cn/models), while international users can access [International Site Online Experience Link](https://cloud.siliconflow.com/models).
In terms of technical features, Kimi K2 has three main advantages: first, large-scale training, which was trained on 15.5T Tokens of data, ensuring the model's stability; second, Kimi K2 uses the MuonClip optimizer, which maintains stability during large-scale expansion; finally, Kimi K2 is a model driven by intelligent agents, designed for tool calling, logical reasoning, and autonomous problem solving.
According to publicly available benchmark test data, Kimi K2 performs well in abilities such as autonomous programming, tool calling, and mathematical reasoning. Moonshot AI stated that Kimi K2 lays a solid foundation for building general agent capabilities, and future updates will include thinking and visual understanding capabilities.
User reviews of Kimi K2 are generally positive. For example, Aravind Srinivas, founder of Perplexity, praised its excellent performance in internal evaluations and said he would soon conduct follow-up training based on this model. Additionally, some users reported that Kimi K2 demonstrated remarkable ability in completing tasks on the web version of Minecraft in one go, far exceeding other models.
Aside from its outstanding coding ability, Kimi K2 has also received praise for its language capabilities, showing a more diverse vocabulary. Users are satisfied with Kimi K2's overall performance and look forward to future updates and expansions.
Key Points:
🌟 The Kimi K2 model has 1T parameters and supports a context length of 128K, performing excellently.
💰 New users can enjoy 14 yuan of trial credit, making it easy to experience the model's functions.
🛠️ Kimi K2 has large-scale training, stable optimization, and an intelligent agent-driven design, suitable for various application scenarios.