Unsloth

1. What is Unsloth?

Unsloth is free, open-source software for running and training AI models on your own computer. Unsloth was once best known as a code library for fine-tuning language models faster and with less memory. The project has since grown into three forms: Unsloth Desktop (a desktop app with a graphical interface), Unsloth Studio (a web interface), and Unsloth Core (the code-based version for developers). You can download a model, chat with it, generate images, or train a model on your own data — all running locally, with no data sent outside.

To help you pick correctly: the "chat with a model" part is beginner-friendly thanks to the click-and-run Desktop app. But training is an advanced task that needs technical knowledge and usually a powerful NVIDIA graphics card (GPU).

A good fit if:

• You want to run an AI model (a chatbot) on your own machine, independent of online services, keeping your data private.
• You are curious to try training a small model and have a strong enough machine.

Not the right fit if:

• You just need a quick chatbot and do not mind using an online service.
• Your machine is weak with no dedicated GPU: you can still chat with small models, but training is essentially not feasible.

2. What you need

For a beginner, use Unsloth Desktop. The notes below focus on that.

• A computer running Windows, macOS or Linux. The official docs list these requirements for Unsloth Studio, the web version, and give no separate minimum for the Desktop app: Windows 10/11 (64-bit), macOS 12 Monterey or newer (Intel or Apple Silicon), Ubuntu 20.04 or a similar 64-bit distribution.
• Free disk space to hold the models. Small models are a few GB, large ones tens of GB, so leave plenty of room.
• No account and no API key are needed for basic use. The software collects no usage data and can run entirely offline (you only need the internet to download models).
• To train (advanced): you need an NVIDIA GPU (Unsloth also supports AMD, Intel GPUs and Mac chips for MLX). The code-based Core version needs Python (version 3.11 to 3.13) and fully installed GPU drivers.

3. Step-by-step installation

Step 1 (easiest): Install Unsloth Desktop on Windows

Go to the official download page https://unsloth.ai/download and choose the Windows build, or download the installer directly from the repo's Releases page. Open the downloaded .exe, follow the setup steps, then launch the app and wait for it to finish its first-time setup.

What you should see: the Unsloth app opens on the Chat page, with a "Select model" box at the top.

Step 2: Install Unsloth Desktop on macOS

Download the Mac build at https://unsloth.ai/download. Open the .dmg file, drag Unsloth into your Applications folder, then launch it and wait for setup to complete.

Step 3: Install Unsloth Desktop on Linux

Download the Linux build (.deb or .AppImage) at https://unsloth.ai/download. For the .deb, open it and choose Install; for the AppImage, make it executable and run it.

Step 4 (advanced): Install Unsloth Studio (web interface)

If you want the browser-based version, open a command window and run the install command for your operating system.

On macOS / Linux / WSL:

curl -fsSL https://unsloth.ai/install.sh | sh

On Windows (PowerShell):

irm https://unsloth.ai/install.ps1 | iex

Once installed, start Studio:

unsloth studio -p 8888

Then open http://127.0.0.1:8888 in a browser. On first launch it asks you to set a password, then opens the Chat page.

Step 5 (for developers): Install Unsloth Core

This is the library for writing training code. If your machine already has Python and PyTorch working with the GPU:

pip install unsloth

See the detailed, per-OS instructions at https://unsloth.ai/docs.

4. First run

Example: download a model and chat with it using Unsloth Desktop.

Step 1: Choose and download a model

Click the "Select model" box at the top, or open the "Model hub" tab. Pick a model and a compression level (quantization) that suits your machine's capacity — smaller models run lighter. Click download and wait for it to finish.

What you should see: the download progress bar reaches 100% and the model appears in your list of installed models.

Step 2: Chat

Type a question into the input box and press Enter. The model answers locally, on your machine.

What you should see: the model answers your question; on a weak machine, the answer may come more slowly.

Step 3: Explore further (optional)

From here you can connect tools (such as Claude Code, web search, MCP), generate images, or start learning about training. Training is much more involved and is best started from the free sample "notebooks" in the official documentation.

5. Common problems and fixes

• nvidia-smi not found (the GPU is not detected) during training. Cause: the NVIDIA GPU driver is not installed. Fix: install the driver from https://www.nvidia.com/en-us/drivers/ and try again.
• nvcc not found when installing Core. Cause: the CUDA toolkit is missing. Fix (Linux): sudo apt install nvidia-cuda-toolkit, or add the CUDA path to PATH.
• After install the model runs very slowly or reports out of memory. Cause: the model is too large for your machine. Fix: choose a smaller model or a lower compression level (for example the 4-bit build) in the Model hub.
• Clicking the button to open Unsloth Studio gives a browser error. A cause noted in the docs: sometimes an ad blocker or disabled cookies. Fix: open the address http://127.0.0.1:8888 directly in your browser.
• A previously downloaded model (for example from Hugging Face) does not appear. Fix: Unsloth usually detects existing models automatically; if not, point it to your model folder in the settings.
• llama-server build failed when installing Studio. Cause: a missing build tool. This is not fatal (Unsloth still runs, only GGUF-model inference is unavailable). Fix: install cmake and rerun setup.

6. Uninstalling and updating

• Update the Desktop app: download the latest build from https://unsloth.ai/download and install it over the old one.
• Update Studio: rerun the same install command (curl … | sh on Mac/Linux, or irm … | iex on Windows).
• Uninstall the Desktop app on Windows: go to Settings → Apps → Installed apps, find Unsloth and remove it.
• Uninstall the Desktop app on macOS: drag Unsloth from Applications to the Trash.
• Delete models to reclaim space: remove them in the app (the bin icon in model search), or delete the Hugging Face model folder at ~/.cache/huggingface/hub/.

7. Frequently asked questions

• Does it cost anything? No. Unsloth is free and open source (Apache-2.0 license; the Studio interface uses the AGPL-3.0 license). You pay nothing to run or train models locally.
• Do I need an internet connection? Yes, to install and to download models. Once a model is on your machine, you can chat entirely offline.
• Is my data sent anywhere? According to the developer, Unsloth collects no usage data; it only detects your GPU/device type to know what will run. The app can run 100% offline.
• Can I use it on a weak machine with no dedicated GPU? You can still chat with small models (even on CPU only), just more slowly; training needs a GPU and is essentially not workable on a weak machine.
• What is the latest version? At the time of writing, the latest Desktop build is v0.1.900-beta (28 September 2026). See the official download page and docs: https://unsloth.ai/download and https://unsloth.ai/docs.