Kimi K3

Kimi K3 is a flagship large-scale language model developed by Moonshot AI, released on July 16, 2026. With 2.8 trillion parameters, it is categorized as the first open-weights model in the 3-trillion-parameter class. The model is designed for frontier intelligence tasks, including long-horizon coding, complex knowledge work, and deep reasoning, natively supporting a 1-million-token context window.

The model's architecture utilizes a Mixture of Experts (MoE) framework, specifically activating 16 out of 896 experts per token to balance performance with computational efficiency. It introduces Kimi Delta Attention (KDA) and Attention Residuals, which are architectural refinements aimed at improving information flow through deep networks and long sequences. These updates, combined with the Stable LatentMoE framework, reportedly provide significant scaling efficiency improvements over previous generations.

Kimi K3 is natively multimodal, supporting the processing of text, images, and video without the need for external adapters. It features an "always-on reasoning" capability where the model performs internal chain-of-thought processing before generating a final response. In technical evaluations and blind developer testing, the model has demonstrated high proficiency in frontend coding and agentic workflows, positioning it near the performance levels of leading proprietary systems.

Official guidance suggests that Kimi K3 is particularly effective at navigating large software repositories and executing complex, multi-step instructions. While the model became accessible via API and integrated web products at launch, the full open-weight release is scheduled to follow in the weeks after the initial announcement.