Install

Install llm-speed

llm-speed is a free, open-source CLI that measures how fast a local LLM runs on your own GPU or Mac: decode tok/s, prefill tok/s and time to first token. Install it, run llm-speed doctor, then llm-speed bench.

Install options

Bare machine (no Python yet)

$ curl -fsSL https://llm-speed.com/install.sh | sh

Sets up Python through uv (it asks first), installs the CLI, then runs llm-speed doctor. The script is served over HTTPS and checks the uv installer's sha256 before running it; read it first at /install.sh if you prefer.

pipx (recommended if you already have Python 3.10+)

$ pipx install llm-speed

Installs llm-speed from PyPI into its own isolated environment.

uv

$ uv tool install llm-speed

Same PyPI package, installed as a uv tool.

Homebrew (tested on Apple Silicon)

$ brew install meadow-kun/tap/llm-speed

From the meadow-kun tap.

After installing

Run the doctor once. It checks every dependency and, in an interactive terminal, walks you through installing whatever is missing (a backend such as llama.cpp, Ollama, MLX or vLLM, and a model).

$ llm-speed doctor

Then measure, and compare your numbers with the cheatsheet of measured runs.

$ llm-speed bench

What gets sent with a run is described on the privacy page; how numbers are measured is on the methodology page.