Why Rapid-MLX Is the Fastest Local AI Engine for Apple Silicon (4.2× Faster Than Ollama)

If your Mac runs Apple Silicon (M1/M2/M3/M4) and you are still using Ollama for local models, you may be missing out on significant speed gains.

How the 4.2× Speedup Is Achieved

Rapid-MLX builds on Apple’s MLX framework, which is specifically optimized for Apple Silicon, and adds model caching, request separation, and inference caching. It is not merely a wrapper around Ollama; it is a ground‑up optimization for the hardware.

Benchmark Results

16 GB MacBook Air – Qwen3.5‑4B – 130 tok/s

24 GB MacBook Pro – Qwen3.5‑9B – 100 tok/s

32 GB+ Mac Mini/Studio – Gemma 4 12B – 42 tok/s

48 GB+ Mac Mini/Studio – Qwen3.5‑35B‑A3B 8bit – 59 tok/s

96 GB+ Mac Studio – Qwen3.5‑122B – 57 tok/s

128 GB+ Mac Studio Ultra – DeepSeek V4 Flash – 31‑56 tok/s

Installation Simplicity

Three convenient methods are provided:

# Homebrew one‑click install
brew install raullenchai/rapid-mlx/rapid-mlx
# Or via pip
pip install rapid-mlx
# Or a single‑line script
curl -fsSL https://raullenchai.github.io/Rapid-MLX/install.sh | bash

After installation, run the engine with a single command: rapid-mlx chat The first run automatically downloads the default model (qwen3.5‑4b). On an M3 chip, the 4 B model reaches 130 tok/s.

OpenAI Compatibility

Rapid-MLX exposes an OpenAI‑compatible API, allowing existing tools (e.g., Cursor, Claude Code, Aider) to point to http://localhost:8000 without code changes.

Key Features

Tool Calling : 17 parsers, including PydanticAI and LangChain structured output.

Prompt Cache : Repeated prompts return instantly, reducing TTFT to 0.08 s.

Reasoning Separation : Separates the model’s “thinking process” from the final answer using FastChat’s conversation template.

For vision models such as Gemma 4, install the vision extension:

pip install 'rapid-mlx[vision]'

Comparison with Other Solutions

Speed : Rapid‑MLX is the fastest (🥇), Ollama is the baseline (1×), LM Studio is close to Ollama.

OpenAI Compatibility : Rapid‑MLX offers full compatibility, while Ollama and LM Studio provide only basic support.

Tool Calling : Rapid‑MLX supports 17 parsers; the others are limited.

Prompt Cache : Available only in Rapid‑MLX.

Reasoning Separation : Available only in Rapid‑MLX.

Model Download : Automatic in Rapid‑MLX and Ollama; LM Studio uses a GUI.

Installation Size : Rapid‑MLX ~460 MB (text only) vs. Ollama ~1.5 GB vs. LM Studio >2 GB.

Rapid‑MLX is currently one of the fastest ways to run local models on Apple Silicon, making it suitable for developers who want to use tools like Cursor or Claude Code, or simply experiment with large models locally.

GitHub: https://github.com/raullenchai/Rapid-MLX<br/> Stars: 2,677 | Language: Python | License: Apache‑2.0