@meta
  v: 1
  route: /runtime/mlx
  generated: 2026-06-10T09:20:50.292Z

@intent
  purpose:    Install, configure, and serve LLMs with MLX.
  audience:   self-hoster, ai-engineer, mac-user, devops-engineer
  capability: install, serve_local_llms, compare_runtimes, open_external_docs

@state
  slug: mlx
  name: MLX
  family: apple-silicon
  type: framework
  license: MIT
  primary_platform: macOS · Apple Silicon
  platforms[3]: macOS (Apple Silicon), Linux (CUDA), Linux (CPU)
  model_formats[1]: MLX
  api_compatibility[4]: "Python (mlx-lm, mlx-vlm)", Swift, C++, C
  install_command: pip install mlx
  install_secondary: pip install mlx[cuda]   # Linux NVIDIA backend
  homepage_url: https://github.com/ml-explore/mlx
  github_url: https://github.com/ml-explore/mlx
  docs_url: https://ml-explore.github.io/mlx/build/html/index.html
  feature_count: 5
  feature_labels[5]: Unified memory, Familiar APIs, Lazy + dynamic, mlx-lm + mlx-vlm, Function transforms
  best_for[4]: Apple Silicon inference (the unified-memory advantage), Researchers extending the framework with custom ops, Building higher-level tools (oMLX is built on this), Training and fine-tuning on Apple Silicon (LoRA on a MacBook)
  caveats[3]: "Not a server. You wrap it in mlx-lm, oMLX, or your own code", Apple Silicon native; Linux paths exist but are not the primary target, The model zoo lives at huggingface.co/mlx-community (community-maintained)

@actions
  - id: open_homepage
    method: GET
    href: https://github.com/ml-explore/mlx
  - id: open_github
    method: GET
    href: https://github.com/ml-explore/mlx
  - id: open_docs
    method: GET
    href: https://ml-explore.github.io/mlx/build/html/index.html
  - id: view_index
    method: GET
    href: /runtime/
  - id: view_calculator
    method: GET
    href: /#calculator

@context
  > Apple's machine learning research framework, designed for Apple Silicon's unified memory architecture. NumPy-style Python API; full C++, C, and Swift bindings. Lazy evaluation, dynamic graphs, composable function transformations. Not a serving runtime on its own; you reach for `mlx-lm` (LLMs) or `mlx-vlm` (vision-language) or build on top with oMLX. Apache 2.0, actively developed by Apple ML Research.

@nav
  self:      /runtime/mlx
  parents:   [/, /runtime/]
  peers:     [/runtime/ollama, /runtime/lm-studio, /runtime/vllm, /runtime/omlx]
  drilldown: /#calculator