Unsloth

Free

A Python library for fine-tuning LLMs [#opensource](https://github.com/unslothai/unsloth).

FreeFree tier

Inputs: text, image, audio, code

Type

Open Source

Company

Unsloth

Visit Website GitHub (40 stars)

About Unsloth

Unsloth is an open-source Python library designed to accelerate fine-tuning of large language models (LLMs) while significantly reducing memory usage. It achieves up to 2x faster fine-tuning on a single NVIDIA GPU with 90% less memory compared to standard Flash Attention 2 (FA2). Unsloth supports over 500 models including Mistral, Gemma, Llama, and more, covering text, vision, audio, and embedding tasks. The library integrates with popular formats like LoRA, FP8, FFT, and PT. Alongside the core library, Unsloth offers Unsloth Studio for running models locally (100% offline on Mac/Windows) with GGUF and Safetensors, tool-calling, web search, and an OpenAI-compatible API. Data Recipes auto-creates datasets from PDF, CSV, or JSON files via graph-node workflows. Models can be exported to safetensors or GGUF for use with llama.cpp, vLLM, Ollama, and others. Free open-source version available; Pro and Enterprise tiers offer additional speed, multi-GPU support, and accuracy improvements.

Key Features

2x faster fine-tuning on a single NVIDIA GPU compared to FA2

90% less memory usage than Flash Attention 2 (FA2)

Supports 500+ models including Mistral, Gemma, Llama 1/2/3, and more

Compatible with LoRA, FP8, FFT, and PT training methods

Covers text, vision, audio, and embedding modalities

No-code training with auto-created datasets from PDF, CSV, and JSON files

Export models to safetensors or GGUF for use with llama.cpp, vLLM, Ollama

Unsloth Studio for 100% offline local inference on Mac/Windows with GGUF and Safetensors

Unlimited tool calling and web search with sandboxed code execution

Data Recipes for transforming unstructured documents into structured datasets via graph-node workflows

Pros & Cons

Pros

Extremely fast fine-tuning and reduced memory usage
Wide model support (500+) covering multiple modalities
Free open-source version available with full core functionality
Local offline runtime via Unsloth Studio
Easy dataset creation from various file types without manual labeling
Export compatibility with major inference engines

Cons

Advanced performance (3x+ speed boost, multi-GPU, multi-node) requires paid Pro or Enterprise plans
Optimal performance tailored for NVIDIA GPUs; AMD support mentioned but limited
Free version limited to single GPU training; multi-GPU only in paid tiers
Some features (e.g., faster inference, full training support) are exclusive to Enterprise plan

Best For

Fine-tuning open-source LLMs for custom tasks on a single GPURunning large language models locally with full offline capabilityCreating training datasets from raw PDF, CSV, or JSON documentsExporting fine-tuned models to deployment formats like GGUF and safetensorsBuilding AI applications with tool-calling and web search functionalityTraining vision, audio, and embedding models with memory optimization

FAQ

Is Unsloth free to use?

Yes, Unsloth offers a free open-source version that supports fine-tuning on a single NVIDIA GPU with 2x speed and 90% memory reduction. Pro and Enterprise plans provide additional speed, multi-GPU support, and accuracy gains.

What models does Unsloth support?

Unsloth supports over 500 models including Mistral, Gemma, Llama 1, 2, and 3, as well as recent models like GLM-5.2, DeepSeek-V4, and DiffusionGemma. It covers text, vision, audio, and embedding models.

Can I run Unsloth offline?

Yes, Unsloth Studio runs 100% offline on Mac and Windows devices, supporting GGUF and Safetensors models with tool-calling, web search, and an OpenAI-compatible API.

How do I export fine-tuned models from Unsloth?

Unsloth supports exporting fine-tuned models to safetensors or GGUF formats, which can then be used with llama.cpp, vLLM, Ollama, and other inference engines.