# MLX

> **TL;DR.** MLX is Apple's open-source array framework for machine learning on Apple Silicon.

- **Category:** AI / Developer Tools / Infrastructure
- **Stage:** established
- **Age:** 924 days
- **Origin date:** 2023-12-05
- **First detected:** 2026-04-20
- **Canonical URL:** https://earlyterms.com/term/mlx
- **Sources:** 8 primary URLs

## Definition

MLX is Apple's open-source array framework for machine learning on Apple Silicon. Its API mirrors NumPy and PyTorch, but the whole runtime is built on Metal and the unified memory architecture, so operations move between CPU, GPU, and Neural Engine without copying tensors.

Open-sourced by Apple's ML Research team in [December 2023](https://github.com/ml-explore/mlx), MLX sat quiet for two years as a research tool. It went mainstream in spring 2026: [Ollama switched its default Apple Silicon backend to MLX on March 30, 2026](https://ollama.com/blog/mlx) (1.6x prefill, 2x decode), the [M5 Neural Accelerators](https://machinelearning.apple.com/research/exploring-llms-mlx-m5) hit 4x faster time-to-first-token, and indie engines like Rapid-MLX now outrun llama.cpp on 16 of 18 benchmarks.

## Example

You install `mlx-lm` with pip, run `mlx_lm.generate --model mlx-community/Qwen3-30B-4bit` on a 32GB MacBook Pro, and get a local 30B model decoding at 60+ tokens/sec — no CUDA, no Docker, no external GPU. Ollama 0.19 ships the same path behind the scenes.

## Analogy

MLX is to Apple Silicon what CUDA is to Nvidia — the native dialect that unlocks the metal underneath.

## Why it's emerging now

Three compounding 2026 events tipped MLX from Apple research project to default Mac inference stack: M5 Neural Accelerators (Oct 2025), Ollama adopting MLX as its Apple Silicon backend (Mar 30, 2026), and a third wave of drop-in engines like Rapid-MLX beating llama.cpp on 16 of 18 models.

## Related terms

- *competitor:* llama.cpp
- *competitor:* Ollama
- *competitor:* MPS (Metal Performance Shaders)
- *competitor:* CUDA
- *related:* Core ML
- *child:* mlx-lm
- *child:* Rapid-MLX
- *child:* mlx-vlm
- *child:* mlx-swift
- *parent:* Apple Silicon
- *related:* lm-studio
- *related:* qwen3

## Sources

1. [GitHub — ml-explore/mlx](https://github.com/ml-explore/mlx)
2. [GitHub — ml-explore/mlx-lm](https://github.com/ml-explore/mlx-lm)
3. [Ollama Blog — MLX preview](https://ollama.com/blog/mlx)
4. [Apple ML Research — LLMs with MLX on M5](https://machinelearning.apple.com/research/exploring-llms-mlx-m5)
5. [The New Stack — Ollama taps Apple's MLX](https://thenewstack.io/ollama-taps-apples-mlx/)
6. [MacRumors — Ollama now runs faster on Macs thanks to MLX](https://www.macrumors.com/2026/03/31/ollama-now-runs-faster-apple-silicon-macs/)
7. [9to5Mac — Ollama adopts MLX for faster AI on Apple Silicon](https://9to5mac.com/2026/03/31/ollama-adopts-mlx-for-faster-ai-performance-on-apple-silicon-macs/)
8. [GitHub — raullenchai/Rapid-MLX](https://github.com/raullenchai/Rapid-MLX)

---
_Generated by EarlyTerms · https://earlyterms.com/term/mlx_