Moonshot AI has released the full weights and technical report for Kimi K3, its most advanced artificial intelligence model. The initiative expands the technical community's access to the system's architecture.
The weights were published on Hugging Face, while the technical documentation was added to the company's official repository on GitHub. Code and weights are available under the Kimi K3 license.
The release fulfills the timeline announced by Moonshot AI in early July. At the time, the company had stated that the complete files and additional specifications would be released by 27 de julho.
Kimi K3 has 2,8 trillion parameters, although only 104 billion are activated during inference. This feature seeks to balance computational capacity and efficiency in task processing.
The model uses a mixture of experts architecture, known as Mixture of Experts (MoE). There are 896 experts available, with 16 selected to process each token during execution.
Another highlight is the context window of more than 1 milhão de tokens. This feature makes it possible to work with large volumes of information in a single interaction, expanding the possibilities for artificial intelligence applications.
The technical report also presents details about Kimi Delta Attention, Attention Residuals, and the native quantization system. According to Moonshot AI, these architectural changes increase scalability efficiency by approximately 2,5 times compared with Kimi K2.
With the files made public, developers can download and implement Kimi K3 using tools such as Transformers, vLLM, and SGLang. The availability facilitates testing, research, and customized applications.
Because of the model's size, Moonshot AI recommends the use of configurations with large-scale supernodes. The required infrastructure is one of the main technical points for teams interested in running Kimi K3 locally.
The release of the weights also allows researchers to directly analyze the model's structure, while developers can assess its performance in different computing environments and artificial intelligence applications.
