Ran Moonshot's 2.8T-parameter Kimi K3 on a GPU-less mini-PC (github.com)

Moonshot AI has announced the release of Kimi K3, a groundbreaking 2.8-trillion-parameter Mixture-of-Experts (MoE) model, which ranks among the largest open-weight models available to the public. Launched on July 27, 2026, Kimi K3 employs a unique architecture that allows it to run on a standard mini-PC without requiring GPUs. The model leverages a Rust engine to optimize performance, utilizing efficient memory management—where the main computation components reside in RAM while the necessary routed experts are streamed from disk on demand. This introduces significant cost and accessibility advantages for developers and researchers working with large-scale AI models. The technical significance of Kimi K3 lies in its innovative approach to resource management and its advanced architectural features. It incorporates Kimi Delta Attention, a Gated MLA hybrid structure, and the Stable LatentMoE routing technique, which improves efficiency by activating only a small fraction of its parameters for each token input. Kimi K3 is designed to streamline processing through advanced quantization (OCP MXFP4) and caching strategies, enhancing speed and reducing disk I/O load. Initial performance metrics indicate a strong foundation, with expectations for optimization to follow, building on previous successes like the GLM-5.2 model. This release marks a significant milestone in making complex AI models more deployable on limited hardware, catalyzing innovation within the AI/ML community.