Rizz on Apple Silicon

Rizz is an inference engine for open weight models. This page covers the native Metal and MLX backend for M series Macs and the command line tools that come with it.

The tools

• rizz serve: start the API server, OpenAI and Anthropic compatible
• rizz shell: chat with a server in the terminal
• rizz code: local coding harness with workspace tools and sessions
• rizz write: cursor aware editing on macOS
• rizz ctl: query and manage a running server
• rizz bench: bandwidth, runtime and kernel benchmarks

Measured

On an Apple M4 Max with 36 GB of unified memory, Qwen3 4B in native MLX Q3 affine decodes at 126 to 129 tokens per second on the Rizz Metal runtime, against about 105 on Ollama for the same checkpoint.

The Rizz engine is open source and was not written at Quilio. It comes from the FlashML project.

Read about Rizz Write · Upstream repository