lmstudio-ai/mlx-engine is the open-source MLX inference engine that LM Studio for Mac uses. It runs on Apple's MLX (a library for running machine learning models on the Mac), is written in Python, and follows the MIT license. It is a project aimed at LM Studio for Mac, and it has 1190 GitHub stars. The repository was created on October 7, 2024.
The engine combines mlx-lm, mlx-vlm, and Outlines. mlx-lm and mlx-vlm handle text generation and vision model inference (models that take an image as input and judge its content), and the structured output component built on Outlines makes the model produce data in a specified format instead of free-form sentences. That is the part needed for tasks that must follow a fixed schema. It ships a standalone demo, so the engine can be run on its own without LM Studio, and speculative decoding (a method that produces candidates first and then has the model verify and confirm them) can also be checked in that demo. This engine is preinstalled in LM Studio for Mac. Text generation, vision inference, and structured output are each delegated to a different library, which matters because the internals of the engine LM Studio uses are open to the public.
What remains on record is limited, though. The momentum score in the research record is a low 1.7, and the novelty, technology, and quality scores were not given. There are no benchmarks comparing performance and no tagged release either, so there is no way to know which models this engine is fast on, or how fast it is. There are no numbers yet to compare its performance against other engines either. It is still too early to use as a basis for performance comparison. For people using LM Studio on a Mac, the repository is worth reading as a reference case for how the inference engine is put together, but it is no ground for an adoption decision. Whether comparison data comes out together with a version release in the future is the next thing to watch.