Make any model yours.

Temper runs, fine-tunes and ships open-weight AI models on your own computer. Your data stays with you.

For macOS, Windows and Linux. Free and open source.

Temper's Train page: a loss chart falling over 40 steps while fine-tuning Qwen3 4B on math word problems.

Chat

Talk to open models. Offline.

Find Qwen, Llama, Gemma, Mistral and thousands more on Hugging Face. Temper checks which quantization fits in your memory, downloads it in the background, and runs it with llama.cpp on Apple Silicon, NVIDIA and AMD.

A chat with a fine-tuned model that solves a word problem step by step.

Train

Fine-tune on the computer you already own.

Pick a base model and a dataset, and press Start. On a Mac, Temper trains LoRA adapters with Apple's MLX. On NVIDIA and AMD GPUs it uses PyTorch with LoRA, QLoRA or full fine-tuning. Loss, validation loss and time left update as it runs.

12 minto fine-tune a 4B model for 60 steps on an M1 Pro
29 MBfor the trained LoRA adapter
0rows of your data sent to a server

Recipes

Your documents in. A dataset out.

Point a recipe at PDFs, Word files, web pages or a GitHub repository. A model writes grounded questions and answers, code with linting, or graded replies, row by row. Preview three rows, then generate thousands and train on them in one click.

The Recipes page: columns for a topic, a question and an answer, with generated rows on the right.

Datasets

See your data the way the model will.

Open any Hugging Face dataset or local file, map its columns to chat turns, and read each row as a conversation before you spend an hour training on it. Remove duplicates, filter by length and hold out an evaluation split in the same place.

The Datasets page mapping a question column and an answer column to chat turns, with a preview of each conversation.

Export

Ship it anywhere.

Merge your adapter and export GGUF in the quantization you need, for llama.cpp, Ollama and LM Studio. Chat with the result one click later, or serve it to coding agents through an OpenAI-compatible API.

The Export page with a finished Q4_K_M GGUF export and a button to chat with it.

Everything in one app.

Knows your hardware

Detects Apple Silicon, NVIDIA and AMD GPUs, shows memory in use, and tells you which models fit before you download them.

Sets itself up

The first launch installs Python and the engine for you, about 150 MB for chat or 600 MB with training on a Mac. Nothing to configure.

Pinned and verified

Every tool Temper downloads is pinned to an exact version and checked against its SHA-256 checksum before it runs.

Works with your tools

Ready-made settings for Codex, OpenCode, Continue and Aider, plus Python and curl examples for your own scripts.

Stays awake

Long training runs and exports keep your computer from going to sleep until they finish, then let it rest.

Open source

Apache-2.0. Read the code, file an issue, or build it yourself from the repository.

Get Temper.

Free for macOS, Windows and Linux.

macOS Apple Silicon and Intel Windows x64 and ARM64 Linux AppImage, deb and rpm

Or install from a terminal

macOS and Linux curl -fsSL https://temper.devops-monk.com/install.sh | sh

Windows irm https://temper.devops-monk.com/install.ps1 | iex

Temper is an early preview. Installers are published on GitHub Releases and are not code-signed yet, so your system will ask you to confirm the first launch. Release notes explain how.