Skip to content
PMetal

Getting Started

Get up and running with PMetal in minutes — install, train your first model, and run inference.

PMetal is a complete machine learning platform for Apple Silicon — from low-level Metal GPU kernels and Apple Neural Engine integration to high-level training APIs, a terminal TUI, and a full desktop GUI.

Prerequisites

Section titled “Prerequisites”
  • macOS on Apple Silicon (M1 or later)
  • Xcode Command Line Tools: xcode-select --install

For building from source, also install:

  • Rust 1.86+ via rustup
  • Metal Toolchain: xcodebuild -downloadComponent MetalToolchain
  • CMake: brew install cmake

Quick Install

Section titled “Quick Install”
Terminal window
# Option 1: Prebuilt binary
curl -fsSL https://github.com/Epistates/pmetal/releases/latest/download/pmetal-aarch64-apple-darwin.tar.gz | tar xz
sudo mv pmetal /usr/local/bin/


# Option 2: Install from crates.io
cargo install pmetal

See Installation for all options including optional serving, MCP, source builds, and GUI setup.

Your First Training Run

Section titled “Your First Training Run”

Fine-tune a model with LoRA in one command:

Terminal window
pmetal train \
  --model Qwen/Qwen3-0.6B \
  --dataset train.jsonl \
  --output ./output \
  --lora-r 16 --batch-size 4 --learning-rate 2e-4

PMetal automatically downloads the model from HuggingFace Hub, detects your hardware capabilities, and tunes kernel parameters for your specific chip.

Run Inference

Section titled “Run Inference”

Chat with your fine-tuned model:

Terminal window
pmetal infer \
  --model Qwen/Qwen3-0.6B \
  --lora ./output/lora_weights.safetensors \
  --prompt "Explain quantum entanglement" \
  --chat

Use the SDK

Section titled “Use the SDK”

Integrate PMetal into your own Rust applications through the facade crate and re-exported modules:

use pmetal::data::Tokenizer;
use pmetal::hub::resolve_model_path;
use pmetal::models::DynamicModel;


let model_path = resolve_model_path("Qwen/Qwen3-0.6B", None, None).await?;
let tokenizer = Tokenizer::from_model_dir(&model_path)?;
let model = DynamicModel::load(&model_path)?;

See Rust Facade for complete inference and training examples.

Or from Python:

import pmetal


result = pmetal.finetune(
    "Qwen/Qwen3-0.6B",
    "train.jsonl",
    lora_r=16,
    learning_rate=2e-4,
    epochs=3,
)

Explore Further

Section titled “Explore Further”