Author(s): Chew Loong Nian – AI ENGINEER

Originally published on Towards AI.

I Ran Claude Code on My MacBook With vllm-mlx — It Embarrassed llama.cpp by 87%

I did something this week that I assumed would be a slow, frustrating downgrade: I unplugged Claude Code from Anthropic’s cloud and pointed it at a model running entirely on my own MacBook. No API key. No per-token bill. No data leaving the machine. I expected a toy. Instead I got a server that pushed 525 tokens per second on a small model, scaled to 4.3x aggregate throughput at 16 concurrent requests, and beat llama.cpp — the undisputed king of on-device inference — by up to 87% on Apple’s own Metal backend. The project is called vllm-mlx, and it shouldn’t be this good.

After the intro, the article explains why local LLM serving on a Mac has become practical (unified memory bandwidth and maturing serving/tooling), what vllm-mlx is (an MLX-native inference server that adopts vLLM’s production-serving approach and ports it to Metal), and how the author tested it across multiple 4-bit-quantized text and multimodal models on an M4 Max. It reports results showing vllm-mlx outperforming llama.cpp in single-stream throughput and scaling significantly better under concurrency thanks to continuous batching, and it highlights a major differentiator: multimodal content-based prefix caching that drastically reduces latency for repeated image/video analysis. The author also notes important caveats (the KV-cache/paged-attention story isn’t fully mature yet, and the main benefits are Apple-Silicon-specific), then gives guidance on when to choose vllm-mlx versus llama.cpp or other tools, and concludes with a “verdict” that positions vllm-mlx as a credible, high-throughput local backend for coding agents—especially when you need concurrent requests or multimodal workflows.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI