What it is
Lemon MLX Engine is a C++ inference engine built on MLX. It runs large language models locally on Apple M-series, AMD GPUs (Linux/Windows), and CPU — no Python required at runtime.
The README lists broad architecture coverage: 50+ LLM families (Llama, Qwen, Gemma, Phi, DeepSeek, Mistral, and more) plus a dozen VLM architectures and embedders.
Why it matters
Python is fine for experiments. A local AI server that people actually install needs a native engine that can target more than one vendor GPU and still behave on Apple silicon. This repo is that engine for the Lemonade stack.
How it works
Models run through MLX. The interesting work, for me, was making the hardware paths honest: Metal on Apple, ROCm on AMD, CPU as a fallback, plus decode-path and fused-operator work visible in the commit history career-database attributed to me (ROCm throughput notes, Metal vs CPU selection, GDN fused decode defaults).
Status
Open source under lemonade-sdk/lemon-mlx-engine. I contributed; upstream maintainers own the roadmap.