davide-troiani/moeshift is a C++ inference tool for running large Mixture-of-Experts (MoE) models locally on Apple Silicon. MoE is a model architecture that computes only a subset of several experts. The project targets memory efficiency and processing performance, and is released under the MIT license. New releases or version changes are not recorded in the provided material. So what this introduction can confirm is not the changes of a particular version, but the implementation and scope of use that the repository describes.
The README presents results comparing Strata with memory usage and processing speed on a Mac mini M4 Pro with Qwen 3.8 Flash Next and a specific quantized model. Quantization is a way of representing a model's weights so that they are handled with less memory. The README also contains the procedure for downloading the model files directly, preparing and verifying the expert weights, and then running a local server. This shows that the repository is not simply one that describes the model, but a tool that guides the process of running it on Apple Silicon.
That said, the project is at the alpha stage, and the README states that it changes quickly and may have rough parts. The presented comparison targets the Mac mini M4 Pro and a specific quantized model, so the provided material offers no grounds for assuming that the same memory usage and processing speed will appear on other devices or models. The specific figures from the comparison are also not given in this material. So for now, this is closer to a stage of examining the implementation and usage procedure for local MoE inference on Apple Silicon than to judging any general performance advantage.