Skip to content

Getting started

These pages take you from nothing to a quantized, exported and evaluated model.

  • 1. Installation

    Install the openternary command, let openternary setup pick the right PyTorch for your GPU, and check everything with openternary doctor. Also covers updating and installing from source.

  • 2. Your first run

    Build a tiny model on your CPU and run the whole pipeline in about a minute: inspect, quantize, pack, restore, validate, evaluate and compare. No download and no GPU needed.

  • 3. The full-screen interface

    Run the same commands from the TUI: choose them from a ❯ prompt, preview the plan, and watch jobs with live progress and ETA.

What you need

Minimum Notes
Python 3.11 3.12 is recommended and is what the release is tested with most. Windows ROCm builds exist only for 3.12 and 3.13.
OS Windows, Linux, macOS WSL works.
uv any recent version Needed for openternary setup and update. Use pip if you prefer, but setup still calls uv.
GPU optional NVIDIA (CUDA), AMD (ROCm) or Apple silicon (MPS). Quantizing runs on the CPU; GPUs speed up evaluation and calibration.
Disk a few GB for small models A ternary fake-quant snapshot is the size of the original model. Calibration adds activation caches and checkpoints.

Already know the CLI?

Every command supports --help. The CLI reference lists every flag, and Automation covers --json, progress events and exit codes.