China-based AI research is increasing its influence in global technology competition. A new large language model, Kimi K2, developed by Moonshot AI, backed by Alibaba, has been released as open source. The model has attracted attention with both its architectural structure and initial performance results.
The Chinese have announced another AI model.
Kimi K2 was developed with a Mixture-of-Experts (MoE) architecture with a total of 1 trillion parameters. However, only approximately 32 million of these parameters are active during each token transaction. This structure allows the model to strike a strong balance between efficiency and transaction costs. Within the model, eight of the 384 expert modules and one collaborative expert are simultaneously activated during each transaction. Kimi K2 consists of 61 layers and was trained with a massive dataset of 15.5 trillion tokens.
During the training process, a specialized optimization method called MuonClip was used to prevent imbalances in the attention mechanism. This technique distributes the model’s attention weights more evenly, ensuring performance stability.
Kimi K2 also offers context window support for up to 128,000 tokens. This means it can process approximately 192 pages of text at once. This feature makes it a standout when working with long documents.
The model is available to users through the free Kimi app. Thanks to its open-source nature, developers can integrate it into various projects. Usage costs are significantly lower compared to existing large language models.
The fee is only 15 cents for 1 million input tokens and $2.50 for an output token. These prices represent a significant difference compared to the $75 output token cost of the Claude model, for example.
Initial user feedback has been overwhelmingly positive in the four days since the model’s release. Reviews on social media, in particular, indicate that the model provides high accuracy in coding tasks. MagicPath founder Pietro Schirano stated that Kimi K2 is the first model he has approached using in production since Claude 3.5 Sonnet.
