MLX Boundary

Why this quest stays shallow on MLX

MLX is the Apple-native ML framework — fast, well-designed, and worth deep study. But it's also a fast-moving target: APIs shift, breaking changes are common, and stale tutorials are misleading more often than helpful. Going deep on MLX in this quest would either go out of date quickly or push the quest into Apple-shaped territory that not every reader has the hardware for.

What you do need to know now

MLX exists. It's Apple-native — built for Apple Silicon and fastest there — though it also ships a Linux CUDA backend (mlx[cuda12]) and a Linux CPU build.
Ollama has an MLX engine on Apple Silicon, previewed in v0.19, but it is opt-in per model: you pull an MLX-built tag such as gemma4:31b-mlx, and the preview asked for more than 32 GB of unified memory. Plain tags are GGUF on Metal. Either way most users never write MLX code directly.
For direct MLX use, the entry point is mlx-lm on PyPI. On Apple silicon that is pip install -U mlx mlx-lm; on Linux add the mlx[cuda12] or mlx[cpu] extra (see the block below).
Models live on the mlx-community HuggingFace org.

Where MLX deep-dives belong

A separate, MLX-specific quest. Until that exists, use Ollama for daily local AI on Apple Silicon and reach for mlx-lm directly only when (a) you need a model that's MLX-only, (b) you're building a custom pipeline that benefits from MLX's NumPy-like API, or (c) you're contributing to MLX itself.

EV v3 · 2026-09-18: Two things here, both wrong from the start. This lesson said MLX is Apple Silicon only; it ships a Linux CUDA backend (mlx[cuda12], mlx[cuda13]) and a Linux CPU build, so the install command and the exercise — which offered only the Apple silicon line — now name all three. And it said Ollama on Apple Silicon v0.19+ uses MLX under the hood, which overstates a preview: MLX is opt-in per model through MLX-built tags, and a plain tag still reports format: gguf. The formats lesson carried the same two claims and moved with this one. Everything else stands — MLX still moves fast, and a dedicated MLX quest is still the right place for depth.

Code

Direct MLX usage when you really need it·bash

# Always start with the latest MLX — APIs shift fast.
# Apple silicon:
pip install -U mlx mlx-lm
# Linux + NVIDIA: pip install -U 'mlx[cuda12]' mlx-lm   (cuda13 wants driver 580+)
# Linux, CPU only: pip install -U 'mlx[cpu]' mlx-lm

# Run a model from mlx-community directly
mlx_lm.generate \
  --model mlx-community/Qwen2.5-7B-Instruct-4bit \
  --prompt "Define unified memory in two sentences." \
  --max-tokens 200

# Programmatic API
python3 - <<'PY'
from mlx_lm import load, generate
model, tokenizer = load("mlx-community/Qwen2.5-7B-Instruct-4bit")
print(generate(model, tokenizer, prompt="Hello!", max_tokens=100))
PY

Why this quest stays shallow on MLX

What you do need to know now

Where MLX deep-dives belong

Code

External links

Exercise

Progress

Comments 0