Running the 2.8T Parameter Kimi K3 Model on a MacBook Pro via SSD Streaming
The Deltafin project demonstrates running Moonshot AI's massive 2.8-trillion parameter model, Kimi K3, on a single MacBook Pro by streaming model weights directly from SSD storage. This technique allows a consumer laptop with modest RAM to execute an AI model far larger than its physical memory capacity. This development proves that trillion-parameter models can run locally on consumer hardware without relying on expensive multi-GPU server clusters. It highlights how novel memory streaming architectures can enable offline, private execution of state-of-the-art AI models on edge devices. Deltafin takes advantage of the dynamic routing in Mixture-of-Experts (MoE) architectures by loading only the necessary expert weights from disk into local RAM for each token. While execution speed is heavily constrained by NVMe read bandwidth—ranging from 1 token/s to 1 token per minute depending on the drive array—it produces exact and fully reproducible outputs.
## BACKGROUND
Large language models usually require high-bandwidth VRAM to hold all parameter weights for real-time inference, restricting trillion-parameter models to massive enterprise clusters. Mixture-of-Experts (MoE) models like Kimi K3 break model capacity into specialized sub-networks and activate only a fraction of parameters per token, making it possible to stream inactive experts on demand from SSDs.