How to Fine-Tune LLMs with Unsloth Studio: Complete Step by Step Guide
By Braincuber Team
Published on April 21, 2026
Unsloth Studio is helping make LLM fine-tuning more accessible through a local, browser-based interface that removes most of the setup and coding overhead. It is designed as a no-code web UI that handles model loading, dataset formatting, hyperparameter setup, and live training monitoring. This complete tutorial walks you through the full workflow for fine-tuning an LLM, from preparing your model and data to running the fine-tuning job, testing the fine-tuned model locally, and finally exporting it for later use.
What You'll Learn:
- How to install Unsloth Studio on your machine
- Understanding the four main sections of the Studio interface
- Choosing and loading base models from Hugging Face
- Selecting the right training method (QLoRA, LoRA, or Full Fine-tuning)
- Configuring hyperparameters for optimal results
- Testing the fine-tuned model in the chat interface
- Exporting models in GGUF format for local inference
What Is Unsloth Studio?
Unsloth Studio is a powerful tool that makes LLM fine-tuning accessible to everyone through a local, browser-based interface. It eliminates most of the setup and coding overhead that traditionally comes with fine-tuning large language models. The platform is designed as a no-code web UI that handles model loading, dataset formatting, hyperparameter setup, and live training monitoring.
Unsloth Studio supports local installation across Linux, Windows, macOS, and WSL, and its export tools support formats such as GGUF, Safetensors, and LoRA for downstream use in tools like llama.cpp, Ollama, and vLLM. For many small-to-medium fine-tuning workflows, this process can be completed surprisingly quickly once the environment is ready, making Unsloth Studio a practical option for beginners who want a simpler path from data to a working fine-tuned model.
System Requirements
| Requirement | Details |
|---|---|
| Operating System | Linux, Windows, macOS, WSL |
| Full Training Support | NVIDIA GPUs, Intel GPUs |
| Chat Only | Mac and CPU-only systems |
| Python Version | 3.11 or newer, but lower than 3.14 |
Step 1: Check Your Setup
Before installing Unsloth Studio, make sure your machine supports training. Unsloth Studio can run on Windows, Linux, WSL, and macOS, but training support depends on your hardware.
Verify GPU and System Requirements
For Linux training, you need Ubuntu, an NVIDIA GPU with drivers installed, Git, and Python 3.11+. Full training is supported on NVIDIA and Intel GPUs. On Mac and CPU-only systems, you can still use Chat and Data Recipes.
Step 2: Install Unsloth Studio
To install Unsloth Studio, use the official one-line installer:
curl -fsSL https://unsloth.ai/install.sh | sh
Installation Time
The first installation can take 5 to 10 minutes because some llama.cpp binaries may need to be compiled.
Step 3: Launch Unsloth Studio
Once the installation is complete, you can start Unsloth Studio with the following commands:
source unsloth_studio/bin/activate
unsloth studio -H 0.0.0.0 -p 8910
If you are running it on your local machine, open this link in your browser:
http://localhost:8910
When you open Unsloth Studio for the first time, it will ask you to create a password for your instance. This helps protect access, especially when you are using a public or externally accessible link.
Step 4: Understand the Studio Interface
Once Unsloth Studio opens, the main screen is divided into four core sections: Model, Dataset, Parameters, and Training/Config. Together, these sections cover the full fine-tuning workflow, from choosing a base model to launching training and monitoring progress.
Model
Choose the base model and training method. Enter a local model path, search for a Hugging Face model, choose the fine-tuning method such as QLoRA (4-bit), and optionally add your Hugging Face token.
Dataset
Select or upload your training data. Choose a dataset from Hugging Face, define the subset and split, and upload your own files if needed.
Parameters
Configure the main training settings such as max steps, context length, and learning rate. There are also expandable settings for LoRA and other training hyperparameters.
Training
Start and monitor training. Once the model and dataset are ready, launch the run, track progress, and upload or save the training config.
Step 5: Choose Your Model
In the Model section, you can either search for a model from Hugging Face or select a local model already available on your machine. Unsloth Studio supports direct Hub search, and local models stored in ~/.unsloth/studio/models or your Hugging Face cache can also appear in the list.
Select Qwen3.5-9B as Base Model
For this tutorial, we will use Qwen3.5-9B as our base model. It is a compact but capable model that works well for small-scale fine-tuning and local experimentation.
Step 6: Pick the Training Method
Unsloth Studio offers three training methods:
| Method | Description | VRAM Required |
|---|---|---|
| QLoRA | Uses a 4-bit quantized base model with a LoRA adapter | Least VRAM |
| LoRA | Keeps the base model in full precision and adds a LoRA adapter | More memory |
| Full Fine-tuning | Updates all model weights | Most VRAM |
For this tutorial, we will use QLoRA because it is the most memory-efficient option and is usually the easiest choice for smaller GPU setups.
Step 7: Load Your Dataset
In the Dataset section, you can load data from Hugging Face or upload local files. Unsloth Studio supports automatic dataset formatting and works with common layouts such as Alpaca, ChatML, and ShareGPT. It also includes tools for split selection, column mapping, and dataset preview.
Load the LaTeX OCR Dataset
Choose the Unsloth Latex OCR dataset, select the train split, and leave the evaluation split empty to keep training faster. Keep the format on Auto so Studio can detect the dataset structure automatically.
Dataset Preview
If Studio cannot fully detect the dataset fields, it opens a preview window where you can map columns such as instruction, input, output, or image. You can also click View dataset to inspect your data before training.
Step 8: Set the Core Hyperparameters
In the Parameters section, Unsloth Studio groups the settings into collapsible panels. For this tutorial, we will keep things simple and only make a few small changes.
| Parameter | Value |
|---|---|
| Epochs | 1 |
| Context Length | 2048 |
| Learning Rate | 2e-4 |
| LoRA Rank | 16 |
| LoRA Alpha | 16 |
Step 9: Start Training
Before starting training, it is a good idea to save your settings. This makes it easier to reload the same setup later or share it with your team so they can reproduce the same results.
Launch the Training Job
Once everything is ready, click Start Training. Studio will open the training view where you can follow the logs as the run begins. You can scroll down to view graphs for metrics such as training loss, learning rate, and gradient norm.
After training finishes, Studio gives you the option to export the model or compare it in Chat. Click Compare in Chat to test the fine-tuned model.
Step 10: Test the Model in Chat
It may take a few seconds for the fine-tuned model to load in the chat interface. Once it is ready, you will see the new model name appear in the Select Model section. You can now test your fine-tuned model with real inputs.
Testing Results
Even though this is only a small example dataset, the fine-tuned model handled LaTeX equations well and showed strong performance on the LaTeX OCR task.
Step 11: Export Your Trained Model
After testing the fine-tuned model, go to the Export section. Here, select your training run, choose the checkpoint, and then pick the export format you want. Unsloth supports exporting as a merged model, LoRA only, or GGUF for local inference tools.
Export in GGUF Format
For this tutorial, export the model in GGUF format and choose Q4_K_M quantization. This is a good option for local use because it keeps the model smaller while still working well in tools like Ollama, llama.cpp, LM Studio, and Jan AI.
Before exporting, Studio will ask whether you want to save the model locally or push it to Hugging Face. Save it locally for this tutorial.
Unsloth Studio Key Features
500+ Models Supported
Fine-tune over 500 open-weight models including Llama 4, Gemma 3, Mistral, Phi-4, DeepSeek-V3/R1, and gpt-oss.
Multi-GPU Support
Unsloth has enabled multi-GPU training. Pool VRAM across multiple GPUs to train larger models or drastically speed up fine-tuning.
Data Recipes
Upload raw files like PDFs, CSVs, and DOCX files and use a visual-node workflow to automatically extract and format text into a training-ready dataset.
RL Support
Supports both standard SFT/QLoRA and Reinforcement Learning algorithms like GRPO. Can cut VRAM usage by up to 80% for RL tasks.
Frequently Asked Questions
Is Unsloth Studio free to use?
Yes. Unsloth Studio is an open-source tool. The core Unsloth package is licensed under Apache 2.0, while the Studio UI is licensed under AGPL-3.0. It is completely free to install and run locally.
What base models does Unsloth Studio support?
Unsloth Studio supports over 500 open-weight models including Llama 4, Gemma 3, Mistral, Phi-4, DeepSeek-V3/R1, and gpt-oss. Beyond standard text LLMs, it also supports Vision, Text-to-Speech, and embedding models.
Do I have to manually format my datasets?
No. Unsloth Studio includes a Data Recipes feature where you can upload raw files like PDFs, CSVs, and DOCX files, and use a visual-node workflow to automatically extract and format text into a training-ready dataset.
Does Unsloth Studio support Reinforcement Learning?
Yes. In addition to standard SFT and QLoRA, Unsloth Studio includes highly optimized support for Reinforcement Learning algorithms like GRPO. It can cut VRAM usage by up to 80% for RL tasks.
What export formats are supported?
Unsloth supports exporting as merged model, LoRA only, or GGUF for local inference tools. GGUF works well with Ollama, llama.cpp, LM Studio, and Jan AI.
Need Help with LLM Fine-Tuning?
Our experts can help you configure Unsloth Studio, optimize fine-tuning parameters, and deploy custom models for your specific use cases.