Unsloth
Unsloth Studio is a web UI for training and running open models like Gemma 4, Qwen3.6, DeepSeek, gpt-oss locally.
Last verified:
What is Unsloth?
Unsloth is an open-source framework for running and training large language models (LLMs) locally on your own hardware via an open-source UI. It streamlines local training, inference, data creation, and deployment with support for Ollama, llama.cpp, and vLLM. Unsloth enables users to fine-tune 500+ models including text, vision, TTS, embedding, and RL models approximately 2x faster with ~70% less VRAM while maintaining no accuracy loss.
Key features include the no-code Unsloth Studio web UI for training and running models, automatic dataset creation from PDF/CSV/DOCX files with visual node workflow editing, self-healing tool calling and web search capabilities, auto inference parameter tuning, side-by-side model comparison, and export to GGUF/16-bit safetensor formats. The framework supports full fine-tuning, pre-training, 4-bit QLoRA, 16-bit LoRA, and FP8 training, plus the most efficient reinforcement learning library using 80% less VRAM for GRPO.
Unsloth is designed for developers, researchers, AI engineers, and anyone wanting to fine-tune LLMs locally without writing code. It directly collaborates with teams behind gpt-oss, Qwen3, Llama 4, Mistral, Gemma 1-3, and Phi-4, fixing critical bugs that improved model accuracy. The tool works on MacOS, Linux, WSL, Windows, NVIDIA, Intel, and CPU setups, requiring as little as 3GB VRAM for notebook-based fine-tuning on Colab or Kaggle.
Unsloth pricing
Pricing model: Freemium
Free tier: Open-source freeware with standard Unsloth version, supports Mistral/Gemma/Llama 1-2-3, 4-bit and 16-bit LoRA, 2x speed boost, 60% VRAM reduction, single GPU support, MultiGPU coming soon. Pro tier: 2.5x faster training + 20% less VRAM, 2.5x number of GPUs faster than FA2, 80% VRAM reduction, enhanced MultiGPU support, up to 8 GPUs, for any usecase, contact us for pricing. Enterprise tier: 30x faster training + multi-node support + 30% accuracy boost, 32x number of GPUs faster than FA2, 90% VRAM reduction, 5x faster inference, supports full training, all Pro features plus multi-node support and customer support, contact us for pricing.
Unsloth pros
- 2x faster training with no accuracy loss
- ~70% less VRAM consumption for fine-tuning
- Open-source and free to use locally
- No-code Unsloth Studio web UI available
- Supports 500+ models across text, vision, TTS, embedding, RL
- Auto-create datasets from PDF, CSV, DOCX files
- Self-healing tool calling and web search support
- Export models to GGUF and 16-bit safetensor formats
- Side-by-side model comparison in Studio
- Auto inference parameter tuning
- Live training monitoring with loss and GPU usage tracking
- Supports full fine-tuning, QLoRA, LoRA, and FP8 training
- Most efficient RL library with 80% less VRAM for GRPO
- Direct collaboration with model creators fixing critical bugs
- Runs locally on Mac, Windows, Linux with 3GB VRAM minimum
- Multi-GPU support with enhanced version coming
- Dynamic 4-bit quants with recovered accuracy
- Visual node workflow for dataset editing
- Works on Colab, Kaggle, and local hardware
- Supports 4x longer context fine-tuning
Unsloth cons
- Multi-GPU version described as incomplete/better version coming
- MacOS Docker support not yet available
- Requires local GPU hardware with sufficient VRAM
- Learning curve for hyperparameter tuning
- Automated evaluation tools may not align with criteria
- Full fine-tuning requires significantly more resources
- Multi-node support only in Enterprise plan
- Pro and Enterprise pricing requires contacting sales
- Limited documentation on advanced use cases
- Manual evaluation often necessary for testing
Frequently asked questions about Unsloth
What is fine-tuning and why is it useful?
Fine-tuning customizes an LLM's behavior, enhances and injects domain-specific knowledge, and optimizes performance for specific tasks. By fine-tuning a pre-trained model on a dataset, you can update knowledge with new domain-specific information, customize tone/personality/response style, and improve accuracy for specific use cases. Fine-tuning can replicate all RAG capabilities but not vice versa.
What is the difference between LoRA and QLoRA?
LoRA is a parameter-efficient training method that adds thin matrices A and B to model weights, optimizing only 1% of weights while keeping the original model in 16-bit precision. QLoRA combines LoRA with 4-bit quantization to save 75% memory, making it accessible for very large models with minimal resources. With Unsloth's dynamic 4-bit quants, the accuracy loss for QLoRA compared to LoRA is largely recovered.
What models does Unsloth support?
Unsloth supports 500+ models including text, vision, TTS, embedding, and RL models. It works with Mistral, Gemma, Llama 1-2-3-4, Qwen3, DeepSeek, Phi-4, NVIDIA Nemotron 3 Omni, Google Gemma 4, and many more. Unsloth directly collaborates with model creators and has fixed critical bugs for gpt-oss, Qwen3, Llama 4, Mistral, Gemma 1-3, and Phi-4.
How much VRAM do I need to use Unsloth?
With Unsloth notebooks on Colab or Kaggle, you can fine-tune for free with just 3GB VRAM. The framework achieves ~70% less VRAM consumption for fine-tuning compared to other methods. The Free tier offers 60% VRAM reduction with single GPU, Pro offers 80% reduction with up to 8 GPUs, and Enterprise offers 90% reduction with multi-node support.
Can I fine-tune without writing code?
Yes, Unsloth Studio is a new open-source web UI that allows you to train and run LLMs with no-code. It provides features like model downloading, LoRA-based fine-tuning, GGUF export, tool calling, interactive chat, automatic dataset creation from PDF/CSV/DOCX, visual node workflow editing, and observability with live training monitoring.
What training methods does Unsloth support?
Unsloth supports full fine-tuning (FFT), pre-training, 4-bit QLoRA, 16-bit LoRA, and FP8 training. It also supports reinforcement learning (RL) with GRPO, GSPO, DPO, ORPO, and KTO. For most use cases, standard Supervised Fine-Tuning (SFT) with QLoRA is recommended as the most accessible method.
How do I deploy my fine-tuned model?
After training, you can save the model as a small 100MB LoRA adapter file or push to Hugging Face Hub. For local inference on single devices, convert to GGUF format using llama.cpp for use in Ollama, llama.cpp, or LM Studio. For enterprise or multi-user deployment with FP8/AWQ, use vLLM. Unsloth streamlines deployment with Ollama, llama.cpp, and vLLM.
What is the difference between instruct and base models?
Instruct models (ending in instruct/chat) allow direct fine-tuning using conversational chat templates like ChatML and ShareGPT, requiring less data. Base models use Alpaca, Vicuna formatting and typically require more data. For beginners, starting with instruct models like Llama 3.1 8B is recommended as they're easier to fine-tune with conversational data.
What datasets work best for fine-tuning?
For LLMs, datasets should have 2 columns (question and answer) in a format that can be tokenized. Well-structured question-answer pairs enhance learning and response accuracy. You can synthetically generate data using ChatGPT or local LLMs, use Unsloth's Synthetic Dataset notebook that parses PDFs/videos and auto-cleans data, or for code fine-tuning, dumping code data without structured formatting can still yield significant improvements.
How do I evaluate if my fine-tuned model works?
You can manually evaluate by chatting with the model to see if responses meet your needs. For automatic evaluation, enable it in Unsloth but reduce dataset size or set evaluation_steps=100 to speed it up. You can also use 20% of training data for testing, though automated tools may not perfectly align with your evaluation criteria, so manual evaluation is often necessary.