I Tested Ollama vs oMLX on Apple M5 Max — 4x Faster Prefill Changes Everything

A hands-on benchmark comparing oMLX (a macOS-native MLX inference server) against Ollama on an Apple M5 Max machine. oMLX leverages Apple's MLX framework natively, with persistent KV caching to SSD and continuous batching, achieving up to 4x faster prefill speeds (2,941 vs 731 tokens/sec) in follow-up queries. The comparison runs three rounds using a Qwen 3.5 9B model, showing oMLX consistently outperforms Ollama — especially in prefill rate — because Ollama still relies largely on llama.cpp rather than full MLX optimization. The post also demonstrates oMLX integration with coding tools like Codex CLI.