How to Run Unsloth Studio Locally: Step by Step Guide for Beginners
By Braincuber Team
Published on April 23, 2026
Unsloth Studio is Unsloth's new open-source local web interface for running, training, and exporting open models from one place. Instead of juggling separate tools for inference, fine-tuning, and testing, it brings everything into a single setup you can run on your own machine. This beginner guide walks you through the complete tutorial of setting up the environment, installing Unsloth Studio locally, launching the interface, and testing models directly from the UI.
What You'll Learn:
- What Unsloth Studio is and its key features
- System prerequisites for running Unsloth Studio locally
- How to set up a RunPod GPU environment (A40 machine)
- Install Unsloth Studio with uv and Python dependencies
- Launch and access the Unsloth Studio web interface
- Test the chat interface with downloaded models and web search
What Is Unsloth Studio?
Unsloth Studio is a local web UI for running and training models, with support for GGUF and Safetensors, side-by-side model comparison, automatic inference parameter tuning, tool calling, and an OpenAI-compatible API. It brings a lot of the Unsloth workflow into one place, making local model experimentation much easier for people who want a visual workflow instead of stitching together scripts by hand.
Key Features of Unsloth Studio
Search and Run Models
Search, download, and run models locally from one unified interface without switching tools.
Side-by-Side Comparison
Compare two models side by side to evaluate response quality, speed, and reasoning differences.
Self-Healing Tool Calling
Use built-in web search and self-healing tool calls for more grounded and accurate responses.
Automatic Parameter Tuning
Automatically tune inference settings for optimal performance without manual trial and error.
No-Code Fine-Tuning
Fine-tune models with a no-code workflow directly inside the Studio interface.
Dataset Creation
Create datasets automatically from files such as PDF, CSV, JSON, DOCX, and TXT.
Export Models
Export models to formats like GGUF and 16-bit Safetensors for deployment.
OpenAI-Compatible API
Access OpenAI-compatible APIs for easy integration with existing applications and tools.
Prerequisites for Unsloth Studio
Before starting the installation, make sure your system has the basic tools needed to run Unsloth Studio locally. The setup is quite manageable, but if you want to fine-tune models instead of only using chat, you will need a more complete GPU-ready environment.
| Prerequisite | Requirement | Notes |
|---|---|---|
| Python | Version 3.11 to 3.13 | Required for running Unsloth and dependencies |
| Git | Latest version | Needed to clone and manage the repository |
| CMake | Latest version | Required for building local components during setup |
| NVIDIA GPU | Required for fine-tuning | Basic chat works on CPU, fine-tuning needs GPU |
| NVIDIA Drivers | Latest version | Recommended to avoid CUDA or training issues |
Note on Hardware
Basic local chat is easier to get running on a CPU, but fine-tuning needs the right NVIDIA hardware and updated drivers alongside the standard setup tools. If you lack local GPU resources, consider using a cloud GPU provider like RunPod.
Step 1: Set Up Your RunPod Environment
For this tutorial, we will use a RunPod A40 GPU machine. You can run everything on your local laptop, but you need enough storage space and fast internet for downloading large models and dependencies. Using RunPod as a cloud option is much faster to set up for this workflow.
Create a New A40 Pod
Start by creating a new A40 pod on RunPod and selecting the latest PyTorch 2.8.0 template from the template options.
Configure Pod Storage
Before deploying, edit the pod settings and set the container disk to 50 GB and the volume disk to 100 GB so you have enough space for installation and model downloads.
Expose Required Ports
Expose HTTP port 8888 for Jupyter Notebook and port 8910 for Unsloth Studio access from your browser.
Add Hugging Face Token (Optional)
Adding your Hugging Face token as an environment variable is optional, but it can help speed up model downloads and make access easier for gated models.
Deploy the Pod
Once everything looks right, save the changes, review the pod summary, and deploy the pod. Within a minute, you will see a link to access JupyterLab.
Step 2: Install Unsloth Studio
Once your RunPod pod is running, open the JupyterLab link from the RunPod dashboard, then launch a terminal. The JupyterLab terminal is quick to access, simple to manage, and lets you keep multiple terminals open at the same time.
Upgrade pip and Install uv
First, upgrade pip and install uv, a fast Python package installer and virtual environment manager.
Create Project Folder
Create a new folder for the project and move into it using mkdir Unsloth-Studio && cd Unsloth-Studio.
Create Virtual Environment
Create a virtual environment using uv venv. You should see output confirming the environment was created successfully at .venv.
Activate Virtual Environment
Activate the virtual environment with source .venv/bin/activate so all subsequent packages are installed in isolation.
Install Unsloth
Once the environment is active, install Unsloth using uv pip install unsloth --torch-backend=auto.
Run Studio Setup Command
Finally, run unsloth studio setup. This step handles the heavier setup including installing Node.js, npm, building llama.cpp, and installing extra Python packages.
pip install --upgrade pip && pip install uv
mkdir Unsloth-Studio && cd Unsloth-Studio
uv venv
source .venv/bin/activate
uv pip install unsloth --torch-backend=auto
unsloth studio setup
Setup Time Warning
The unsloth studio setup step can take 15 to 20 minutes to complete. The docs mention 5 to 10 minutes, but in practice it often takes longer as it installs Node.js, npm, builds llama.cpp, and installs extra Python packages. Be patient during this step.
Step 3: Launch Unsloth Studio Locally
Remember that at the start, we exposed port 8910 while editing the RunPod template. We will now launch Unsloth Studio on that port so we can access it directly through the RunPod link instead of only using localhost inside the machine.
Start Unsloth Studio
Run unsloth studio -H 0.0.0.0 -p 8910 to start the web interface on the exposed port with host binding for external access.
Check Detection and Default Model
Once it starts, you should see that it detected the A40 GPU, along with default admin credentials and a default model (usually Qwen3.5 4B quantized) already downloaded and ready to use.
Access the Web Interface
Go back to the RunPod dashboard and open the Connect tab. Click the link for port 8910 to open the Unsloth Studio web interface in your browser.
Create a Password
The first time you open it, Studio will ask you to create a new password. Set one so the instance is protected and not exposed to anyone else using that public link.
Explore the Dashboard
Once inside, you will land on the Unsloth Studio dashboard with four main tabs: Studio for fine-tuning, Recipes for dataset creation, Export for model export, and Chat for model interaction.
Step 4: Test the Unsloth Studio Chat Interface
Click on the Chat tab, then click the Select model button in the top-left corner. The Qwen3.5 4B model is usually already downloaded by default. Click on it, select the installed quantized version, and the model will load into GPU memory within seconds.
| Feature | Description |
|---|---|
| Model Selection | Search and download any model from Hugging Face directly inside the interface |
| Web Search | Enable web search to retrieve recent information and generate grounded responses |
| Chat Configuration | Adjust generation parameters, add system prompts, modify chat templates |
| SVG/Code Generation | Generate SVG art, code, and structured outputs with reasoning capabilities |
You can test the chat interface by giving it math problems, enabling web search for recent topics, or asking it to generate SVG art or code. The responses are fast, smooth, and coherent. You can also download larger models like Qwen3.5 9B 4-bit quantized directly from the interface.
Storage Note
The full setup can take close to 100 GB of storage space. This includes the virtual environment, downloaded dependencies, model files, build artifacts, and tools installed during setup. Make sure your volume disk is sized accordingly (100 GB recommended).
Final Thoughts
Overall, Unsloth Studio offers a polished, fast, and surprisingly easy-to-use interface once it is running. The biggest challenge is the initial setup time and storage requirements, but once Unsloth Studio actually launches, the experience is genuinely enjoyable. Everything is in one place, the workflow feels smooth, and it makes local model experimentation much more approachable.
While the setup may test your patience a little, the actual product experience makes it feel worth it. With a larger and stronger model, you can usually get much better results, especially if your system has a more powerful GPU and enough RAM.
Frequently Asked Questions
What is Unsloth Studio?
Unsloth Studio is an open-source local web interface for running, training, and exporting open models from one place, supporting GGUF, Safetensors, tool calling, and OpenAI-compatible APIs.
Do I need a GPU to run Unsloth Studio?
Basic local chat can run on a CPU, but fine-tuning models requires an NVIDIA GPU with updated drivers and a proper CUDA setup.
How long does the Unsloth Studio setup take?
The setup command can take 15 to 20 minutes in practice, as it installs Node.js, npm, builds llama.cpp, and installs extra Python packages needed for the run.
Can I download models from Hugging Face in Unsloth Studio?
Yes, you can search for any model from Hugging Face and download it directly inside the Unsloth Studio interface with just a few clicks.
How much storage space does Unsloth Studio need?
The full setup can take close to 100 GB including the virtual environment, dependencies, model files, and build artifacts. A 100 GB volume disk is recommended.
Need Help with AI Implementation?
Our experts can help you implement local LLM setups, Unsloth fine-tuning workflows, and custom AI model deployment for your business.