Unsloth

Unsloth can be used in three ways: the Unsloth Desktop app, the Unsloth Studio web UI, or the code-based Unsloth Core.

Verified

Core facts

Read by a script from the official API, not rewritten by a model.

License
Apache-2.0 (core package) / AGPL-3.0 (Studio UI and optional components)
Languages
Python
Maintainer
unslothai
Verified

What it is

Unsloth can be used in three ways: the Unsloth Desktop app, the Unsloth Studio web UI, or the code-based Unsloth Core. Unsloth works on Windows, Linux, WSL and macOS. The project supports multi-GPU setups, NVIDIA, AMD and Intel GPUs, CPUs and the Vulkan backend. Unsloth runs and trains LLMs, MLX, GGUF, diffusion, decision, embedding and audio models.

Main features

  1. Unsloth lets users use local models with Claude Code, Codex and MCP, including tool calling and code execution.

  2. Unsloth offers private web search, deep research, auto-compaction (rolling context window) and RAG.

  3. Unsloth serves models through an OpenAI compatible API.

  4. Unsloth supports reinforcement learning, LoRA, QLoRA, full fine tuning, pretraining, RL, GRPO, DPO and FP8.

  5. Unsloth builds datasets from PDFs, CSVs, DOCX files and more with Data Recipes.

Getting started

  1. The native Unsloth Desktop app can be downloaded for your operating system.

  2. Downloads come from the Unsloth site or from GitHub Releases.

  3. The Unsloth Docker image is available on Docker.

  4. There are four command blocks, in this order: Unsloth Studio (web UI) install for macOS, Linux, WSL; Studio install for Windows; Unsloth Core (code-based) install for Linux, WSL; Core install for Windows. Choose the component first, then the platform, and run only one.

  5. For notebooks (Jupyter or Colab compute notebooks) only, the core tag of that Docker image is used.

  6. The project's notebooks allow training for free.

  7. The notebook workflow is to add a dataset, run, then deploy the trained model.

  8. Unsloth Start connects Claude Code, Codex and other agents to local models with one command.

From the README · bash
curl -fsSL https://unsloth.ai/install.sh | sh
From the README · powershell
irm https://unsloth.ai/install.ps1 | iex
From the README · bash
curl -LsSf https://astral.sh/uv/install.sh | sh
uv venv unsloth_env --python 3.13
source unsloth_env/bin/activate
uv pip install unsloth --torch-backend=auto
From the README · powershell
winget install -e --id Python.Python.3.13
winget install --id=astral-sh.uv  -e
uv venv unsloth_env --python 3.13
.\unsloth_env\Scripts\activate
uv pip install torchao
uv pip install unsloth --torch-backend=auto

Who it is for

The project advises keeping the password safe, or disabling tools, when exposing Unsloth. Exposing Unsloth asks once at the terminal for a new admin password. The Global HTTPS Access option creates a free Cloudflare link that serves Unsloth.

Unsloth uses a dual-licensing model of Apache 2.0 and AGPL-3.0. The core Unsloth package is licensed under Apache 2.0, while certain optional components, such as the Unsloth Studio UI, are licensed under the open-source license AGPL-3.0. The project provides an AMD Guide and an Intel Guide for installing Unsloth on AMD and Intel GPUs.

AMD has a separate Docker image. Old model files can be deleted from the bin icon in model search, or by removing the relevant cached model folder from the default Hugging Face cache directory. The llama.cpp library lets users run and save models with Unsloth.

Sources

This page is written from the official sources below. Open them to check the original text.

Verified ·

This page is written by AI from the official sources above and checked automatically. This site has not tested it. Found a mistake? Tell us.