Skip to content
mlx.app

Handbook

Read in order for a full grounding, or jump to the one that is blocking you. Every article is written for someone who owns a Mac, not someone who owns a datacentre.

  1. 01 · 6 min

    Unified memory, explained properly

    Why a Mac with 48 GB behaves unlike a PC with a 24 GB graphics card, and where the limit actually sits.

  2. 02 · 7 min

    Quantization without the hand-waving

    What 4-bit actually means, why the file is bigger than the arithmetic suggests, and when the quality loss starts to matter.

  3. 03 · 6 min

    The KV cache is the thing that surprises people

    Your model fit at 4K context and died at 128K. This is why, and how to calculate it in advance.

  4. 04 · 6 min

    Why tokens per second is a bandwidth problem

    Decode speed is almost entirely predictable from memory bandwidth and model size. Here is the arithmetic.

  5. 05 · 5 min

    Mixture-of-experts: big model, small bill at inference

    Why a 30B MoE can feel like a 3B model, and what it costs you in memory to get that.

  6. 06 · 5 min

    Context windows: advertised, usable, and affordable

    A model that claims 128K rarely gives you 128K useful tokens, and you probably cannot afford them anyway.

  7. 07 · 4 min

    Your first model in about five minutes

    Install, download, generate. Nothing to configure, nothing to sign up for.

  8. 08 · 6 min

    How to choose a model without reading a leaderboard

    Benchmarks measure benchmarks. Here is a decision procedure based on what you are actually doing.

  9. 09 · 6 min

    Buying a Mac for local AI

    Which spec actually changes your experience, and which upgrades are money you will not notice.

  10. 10 · 5 min

    Quantizing a model yourself

    When the quant you need does not exist, converting it is one command and about ten minutes.

  11. 11 · 6 min

    Fine-tuning on your own Mac

    LoRA on Apple silicon is real and practical. What it can teach a model, and what it cannot.

  12. 12 · 5 min

    What running locally actually protects

    The privacy argument for local inference is strong, but it is narrower than the marketing suggests.

  13. 13 · 6 min

    When it does not work

    The failures people actually hit on Apple silicon, and what fixes each one.