paddock
A command-line tool that tells you which local LLMs fit an Apple Silicon Mac, how fast they will run, and how to launch them, in one command.
paddock answers three questions that picking a local model normally means working out by hand: whether a model fits, given that macOS caps how much memory the GPU can actually use; how fast it will run, since a model that only reaches a few tokens per second is not usable; and how to launch it, replacing manual Ollama tags or llama-server flag combinations.
It reads macOS's real Metal working-set limit directly to budget whether a model fits fully on the GPU, and ships a per-Apple-chip memory-bandwidth table from the M1 through the M5 generation, since Apple Silicon token generation is memory-bandwidth bound and bandwidth varies widely across the lineup. Speed estimates model per-token traffic as active weights plus an 8000-token-deep key-value cache, so they reflect real long-context decoding rather than an empty-context best case.
It treats Apple's own MLX framework as a peer of Ollama and llama.cpp, and running it with no arguments opens an interactive terminal interface that loads instantly from the last catalog snapshot and refreshes in the background.