Moonshot AI released their latest artificial intelligence (AI) model Kimi K2 as open source this morning. The model has a total of 1 trillion parameters with 32 billion parameters active at any one time using a mixture-of-experts (MoE) architecture. MoE is a model that is trained to learn a specific thing separately to become an expert on it.
According to Moonshot, Kimi K2 has high capabilities in boundary knowledge, mathematics and hosting. Two models were launched: Kimi-K2-Base, which is a basic model for researchers and developers who want full control for fine-tuning and personalized solutions. Kimi-K2-Instruct is a model trained for general conversation and agents that answer questions reflexively without thinking.
In several benchmark tests, Kimi K2 managed to score better than the latest models from OpenAI, Google, DeepSeek and Perplexity. The advantage of Kimi K2 is that it is provided for free without having to pay as other companies do. For those who want to use Kimi K2, Moonshot AI offers it through the GitHub page.