AI BriefingKimi-K3: 2.8T Parameter Open-Weight Multimodal Agent
moonshotai/Kimi-K3
About the project
This is a 2.8T parameter open-weight multimodal agent model. By applying the Kimi Delta Attention and Stable LatentMoE architectures, it activates 16 out of 896 experts, achieving approximately 2.5x scaling efficiency compared to Kimi K2. It features an MoE structure with 104B active parameters and supports a 1 million token context window.
It offers native multimodal capabilities, processing text, images, and video within a single model. During long coding sessions, it navigates large repositories, controls terminal tools, and performs complex engineering tasks such as GPU kernel optimization and chip design. It automates knowledge work ranging from research report writing to interactive visualization and motion design.
Deployment efficiency is enhanced through quantization-aware training using MXFP4 weights and MXFP8 activations. It scores 81.2 on the FrontierSWE benchmark, demonstrating high performance compared to competing models. Full model weights are released under the Kimi K3 license, enabling research and commercial deployment.
moonshotai/Kimi-K3
The original page has no description.
image-text-to-text
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.