zevin17/deepinfer-mac is a Python inference framework that runs local models with MLX on Apple Silicon. It is a community port that brings over DeepInfer's scheduling components. The repository description states that it provides local inference and an API, along with reproducible benchmarks. There are no public releases.
The command line, the Python interface, the compatible chat API and the browser screen all use the same runner. The project applies DeepInfer's approach of steering inference with draft and verification schedules to MLX execution. The draft acceptance rate is the share of draft answers that the verification stage accepts as-is. This approach looks at the draft acceptance rate and the overall cost together, and context reuse and state saving are also covered in the local execution flow. The documentation distinguishes between a default path and an adaptive path, and explains execution and state reuse according to model and cache features. The license is Apache-2.0.
The README presents the verified scope per path for Qwen and SmolLM2, and some combinations remain to be checked. The performance records are limited to a specific device and execution conditions, and independent reproduction is not presented. So even though it advertises reproducible benchmarks, it is difficult to judge superiority between devices or general performance from the published performance numbers alone.