vllm-mlx

vllm-mlx

An OpenAI and Anthropic compatible inference server for Apple Silicon, supporting multimodal models.

MCPDevOpen source
Type
MCP
Transport
http
Open source
Yes
GitHub Stars
★ 1.4k
Source
mcp-github

Overview

vllm-mlx is an inference server designed for Apple Silicon, enabling the execution of LLMs and vision-language models (such as Llama, Qwen-VL, LLaVA). It offers continuous batching, MCP tool calling, and multimodal support. Built natively on the MLX backend, it achieves throughput of over 400 tok/s and is compatible with Claude Code. This capability allows AI models to run efficiently on Apple Silicon across various use cases requiring high-performance inference.

Capabilities

  • OpenAI and Anthropic API compatibility
  • Continuous batching
  • Multimodal support
  • High-throughput inference
  • MCP tool calling

Use cases

Multimodal inferenceHigh-performance text generationImage and audio processingStructured output

Setup

Requires: Apple Silicon (M1, M2, M3, M4, M5)Python 3.10+
pip install vllm-mlx

This information was compiled by AI from public sources and may contain inaccuracies — please refer to the source.

FAQ

vllm-mlx supports which hardware?

Only Apple Silicon (M1, M2, M3, M4, M5).

How do I start the vllm-mlx server?

Use the command `vllm-mlx serve <model> --port 8000`.

Related skills