01 · The idea
What stayed with us
MLX is a foundational Mac-local reference for training and inference. We study its implementation and use the platform where it fits while keeping our own model and evaluation contract.
02 · The local translation
What we did with it
MLX gives this project a native Apple Silicon path for tensor work and model adaptation. The learning value is also architectural: its APIs and examples show what the platform handles, letting us spend effort on the data, evaluation, and packaging questions that remain ours.
03 · The boundary
Where the comparison stops
Using MLX does not make every model trainable within one Mac's memory or make a specialist good at its task. Each run still needs a bounded preset, a frozen baseline, and a task-specific gate.