Moonshot AI Opens the Weight of Kimi K3, an AI Model with 2.8 Trillion Parameters

JAKARTA - Moonshot AI has opened the full weight of Kimi K3, an artificial intelligence model with a total of 2.8 trillion parameters. The Beijing-based startup claims Kimi K3 as the world's first open source model in the 3 trillion parameter class.

Yicai Global, quoted Wednesday, July 29, reported that the weight of the model was uploaded to Hugging Face on July 27. Moonshot AI also released a technical report as well as three main technologies for training infrastructure, namely MoonEP, FlashKDA, and AgentEnv.

Kimi K3 is released under a modified MIT license. The license allows developers to download, modify, refine, and run the model locally.

Model weights are a set of values that determine how an AI system processes information and produces answers.

Moonshot AI tested Kimi K3 using Nvidia H200 chips and graphics processing units from other manufacturers. The test showed that the implementation of Kimi K3 is not limited to the Nvidia ecosystem.

Kimi K3 has a total of 2.8 trillion parameters. However, only 104 billion parameters are active when the model performs a certain task.

This model uses the Mixture of Experts architecture. In the system, each token only activates 16 of the 896 special parts called experts.

Tokens are small units processed by the AI model. Their shape can be words, word fragments, numbers, or punctuation marks.

Kimi K3 also uses the Kimi Delta Attention mechanism and Moonshot AI's Attention Residuals technology. The technology supports visual understanding and context windows of up to one million tokens.

Compared to Kimi K2, the scaling efficiency of Kimi K3 is claimed to increase by about 2.5 times.

According to Moonshot AI's technical report, Kimi K3 outperforms all open source models and most closed models in dozens of tests. The tests cover programming, reasoning, visual capabilities, and AI agent workflows that can run a series of tasks.

On Arena Intelligence's WebDev Arena, Kimi K3 scored an Elo score of 1,679. Moonshot AI said the result made it the first open source model to occupy the top position on the ranking board.

Kimi K3 also scored 57.1 on the Artificial Analysis Intelligence Index version 4.1. This result places him in fourth place in the world.

However, the report said that Kimi K3's performance was still slightly behind Claude Fable 5 and GPT-5.6 Sol.

Moonshot AI said the opening of the weight, technical report, and supporting technology of Kimi K3 aims to accelerate the use of cutting-edge AI. This step is also aimed at encouraging research on artificial general intelligence.

The company assessed that the launch marked the entry of the open source large language model into the 2.8 trillion parameter stage.