Kimi K3 AI, a 2.8‑trillion‑parameter model, launches with open‑access performance

The Kimi K3 AI model, released by Chinese developers, features 2.8 trillion parameters and is offered as an open‑access large‑scale language model. It employs novel mechanisms called Kimmy Delta Attention and Attention Residuals, which improve memory efficiency and output precision by up to 2.5×.

Benchmarks reported by Two Minute Papers show Kimi K3 outperforming GPT‑5.6 in the SWE Marathon, and the model is capable of generating fully functional operating systems, designing immersive games comparable to Animal Crossing, and supporting a wide range of scientific, creative, and professional workflows. A GitHub project demonstrates how the model can be run on a single Apple Silicon Mac, streaming expert weights on demand to keep local storage requirements manageable.

The open‑access approach is intended to remove proprietary barriers, allowing researchers, educators, and developers worldwide to experiment with and scale the model for diverse use cases.