Installation

Minimum hardware

If you're not sure whether your machine meets these, run haiku --version after install — it prints a short readiness summary on first launch.

One-line install

curl -fsSL https://haiku.run/install.sh | bash

The installer:

1. Checks that you're on macOS with Apple Silicon (arm64)
2. Ensures the Metal toolchain is available for GPU inference
3. Downloads and verifies the signed, notarized haiku binary
4. Adds ~/.haiku/bin to your shell's PATH
5. Sets up shell completions
6. Downloads the required local AI models (text, vision-language, coder)

Open a new terminal after installation, then verify:

haiku --version

Additional coder models

The installer fetches the three required models — text, vision-language, and the default coder. Two larger coder models are optional (~16 GB each) and can be downloaded later via the CLI:

haiku models install qwen3-moe   # Qwen3-30B-A3B MoE — bigger coder
haiku models install qwen3-6     # Qwen3.6-27B — coder + vision

Run haiku models to see what's installed and what's available. Pass any shorthand to recite / record via -m to use it for a particular run:

haiku recite test_parser.rs.haiku -m qwen3-moe

To reclaim disk space, remove a model:

haiku models remove qwen3-moe

Required models (text, vision, default coder) cannot be removed individually — removing any of them would break recite. To wipe haiku entirely, run haiku uninstall.

Uninstall

haiku uninstall

This removes the binaries, all downloaded models, and the PATH entry from your shell profile. It will ask for confirmation before doing anything.

Telemetry

Haiku sends an anonymous install ping at the start and end of an install, and another when you run haiku uninstall. The pings include your macOS version, CPU model, RAM, free disk, whether Xcode is installed, the Haiku version, and your country — and nothing else. The install ID is a one-way hash of your machine's hardware UUID; the raw UUID never leaves your machine.

To opt out, set HAIKU_NO_TELEMETRY=1 before running install.sh or any Haiku command. Full details, including the hash recipe so you can verify what we receive, are on the privacy page.