MTPLX setup and Qwen3.6-27B speed test on M3 Max

This tutorial tests MTPLX, an Apple Silicon package that uses multi-token prediction and speculative decoding to accelerate local LLM inference. The presenter runs Qwen3.6-27B on a MacBook Pro with an M3 Max. The supplied metadata lists 36GB of unified memory and macOS Tahoe.

The setup starts with a Homebrew installation command from the project's repository. The presenter then shows model selection, including a Q4 quantized option, and explains that higher precision is an option for users who prioritize quality. MTPLX provides terminal chat and a local browser chat interface, with controls for reasoning and MTP.

The reported 192-token test reaches 24.68 tokens per second with MTP enabled. After disabling MTP, the presenter reports 15.35 tokens per second. Those results support a speed improvement in this demonstration, but fall short of the roughly twofold claim in the video's title. The introduction cites a separate 2.24-times decoding speed claim.

During generation, the presenter reports about 15.26GB of RAM usage and a 0.4-second time to first token. These observations describe this model and setup; they do not establish minimum memory requirements or performance across other Macs.