Unsloth Documentation

What Unsloth is

Unsloth is an open source project for running and training open models. It began as a library that makes fine-tuning faster and lighter on memory, and has grown into a desktop app and a set of tools for running models locally too. For the MLOps roadmap the part that matters is fine-tuning: LoRA, QLoRA and full fine-tuning of open LLMs, plus support for reinforcement learning methods such as GRPO and DPO.

Why learners use it

Fine-tuning a model of several billion parameters used to need expensive hardware and a lot of setup. Unsloth's README claims training up to twice as fast with around 70 percent less GPU memory than standard approaches, and it publishes free notebooks that run on Google Colab and Kaggle for many popular open models. That means you can run a real QLoRA fine-tune on a free notebook GPU and see every step: loading a 4-bit base model, attaching adapters, formatting data with the right chat template, training and saving.

What the docs cover

• Installation on Windows, macOS, Linux and in Docker, and the desktop app.
• Guides per model family, with recommended settings.
• Fine-tuning concepts: LoRA rank, learning rate, epochs and how to spot overfitting.
• Datasets and chat templates, which decide more of the result than hyperparameters do.
• Saving and exporting: merged weights, GGUF for llama.cpp and Ollama, and formats for serving with vLLM.

Prerequisites

Python, basic PyTorch ideas, and an understanding of what fine-tuning is for. Know why you are fine-tuning before you start: for most tasks, better prompts or RAG come first.

How long it takes

A first notebook fine-tune takes an afternoon. Preparing a clean dataset of your own and evaluating the result properly against a baseline is a week or two of part-time work, and it is where the real learning happens.

Free or paid?

The core package is free and open source under Apache 2.0, and optional components such as the Studio UI are under AGPL-3.0. The notebooks are free to run on free notebook GPUs, within those services' limits.

Common questions

Do I need my own GPU? No. The free notebooks are enough to learn on small models.

What can I do after it? Fine-tune a small open model on your own task with LoRA or QLoRA, export it for serving and compare it with the base model and an API model on your eval set.

Best for: engineers who want hands-on fine-tuning experience without renting large GPUs.