node-mlx

The fastest way to run LLMs in Node.js on Apple Silicon.

2× faster than alternatives. Powered by Apple MLX.

$npm install node-mlx

Terminal

$ npx node-mlx "What is 2+2?"

✓ Loading Qwen3-4B-Instruct...
⚡ Generated 24 tokens at 142 tok/s

The answer is 4.

Why node-mlx?

2× Faster

Native Metal GPU acceleration. Outperforms node-llama-cpp on every benchmark.

Unified Memory

No CPU↔GPU copying. Apple Silicon's unified architecture = maximum efficiency.

Zero Config

npm install and you're ready. Auto-downloads models from HuggingFace.

Performance

~1.9× faster than node-llama-cpp on Apple Silicon. Benchmarks on Mac Studio M1 Ultra (64GB), 4-bit quantization.

Model

Size

node-mlxllama-cpp

14B

Simple API

app.ts

import { generate } from "node-mlx";

// One-liner: load, generate, done
const result = generate("qwen", "Explain quantum computing:");

console.log(result.text);
console.log(`${result.tokensPerSecond} tok/s`);

Supported Models

Use short aliases or any model from mlx-community

Qwen

Alibaba

4B–32BRecommended

Phi

Microsoft

14BHigh Quality

Gemma

Google

1B–27BLatest

Llama

Mistral

Mistral AI

3B–14BMinistral

Built on

Node.js

Apple MLX

Apple Silicon

Ready to run LLMs at native speed?

Get Started