MLX
Apple's machine learning framework built for Apple silicon.
MLX terms, in plain English
Short, clear definitions for the words that make docs feel heavier than they need to be.
Apple silicon
Apple's M-series chips, like M1, M2, M3, and newer.
Array
A grid of numbers in memory. In ML work, arrays hold inputs, activations, and weights.
Tensor
A general ML word for a multi-dimensional array. Same basic idea, just broader terminology.
Autograd
Automatic differentiation: the system that figures out gradients for training.
Gradient
A signal showing how to change model weights to reduce error.
Lazy evaluation
Work gets delayed until the result is actually needed, which can help performance.
Dynamic graph
The computation graph forms as your code runs, which makes experimentation easier.
Unified memory
CPU and GPU share memory on Apple silicon, reducing expensive copying.
Inference
Running a trained model to get an output.
Fine-tuning
Training an existing model a bit more so it gets better at a specific job.
LoRA
A lightweight fine-tuning method that updates a small adapter instead of the whole model.
Quantization
Shrinking model precision so it uses less memory and often runs faster.
Tokenizer
The tool that breaks text into pieces a model can process.
Weights
The learned numbers inside a model. Loading weights means loading what the model has learned.
Checkpoint
A saved model state you can reload later.
GPU
The graphics processor, often used to speed up ML computation.
Batch
A group of examples processed together for efficiency.
Context window
How much text a language model can consider at once.
Prompt
The instructions or input text you feed into a model.
VLM
A vision-language model: a model that can reason across images and text.
ONNX
A model format and interchange layer used to move models across frameworks or runtimes.
Bindings
A way to use the same core library from another language like Swift or C.
Local inference
Running the model on your own machine instead of calling a cloud API.