Unsloth Docs | Unsloth Documentation

Unsloth is an open-source framework and desktop app for running, fine-tuning, serving, and deploying language, vision, diffusion, embedding, and audio models on local hardware.

Visit Website
Unsloth Docs | Unsloth Documentation

Introduction

Overview

Unsloth is an open-source framework for running and training AI models on local hardware. Its free desktop application brings model discovery, chat, fine-tuning, media generation, serving, and deployment into one interface for macOS, Windows, and Linux.

The platform supports language, vision, diffusion, embedding, and audio workloads. It can work with local models as well as configured cloud providers and OpenAI-compatible endpoints.

Key Features

  • Runs and trains text, image, video, embedding, and audio models.
  • Supports reinforcement learning, LoRA, QLoRA, full fine-tuning, pretraining, GRPO, DPO, and FP8 workflows.
  • Builds datasets from formats including PDF, CSV, DOCX, and JSON.
  • Exports models in formats such as GGUF, NVFP4, and FP8.
  • Serves models through an OpenAI-compatible API, a local network, or secure Cloudflare HTTPS access.
  • Provides web search, deep research, RAG, tool calling, code execution, and MCP connections.

Unsloth states that its fine-tuning approach can train models twice as fast while using 70% less VRAM without accuracy loss. It also reports support for inference and training across more than 500 models.

Getting Started

Unsloth supports NVIDIA, AMD, Intel, and CPU setups across its supported operating systems. In the desktop app, users install the application, choose a model and a quantization suited to their device, download it, and then begin chatting or configuring another workflow.

Permission controls govern whether tool-using models can access files or the internet. Tools can run in a secure environment, while users may separately allow direct file access and editing when needed.

Local Agents and Deployment

Unsloth Start connects local models to coding and agent tools including Claude Code, Codex, DeepSeek Harness, OpenCode, Hermes Agent, and OpenClaw. Users first load a model, open a project folder, and then start the chosen agent through the Unsloth command.

Models can be served over a LAN so traffic remains within the local network, or exposed remotely through secure HTTPS. The API configuration allows control over context size, GPU layers, threading, sampling, networking, and tool behavior.

Training and Media Workflows

The desktop app supports no-code fine-tuning from imported datasets, including LoRA, full fine-tuning, and pretraining. It can also train diffusion LoRA adapters from a user's images and load exported adapters for inference.

For media work, Unsloth can generate and edit images and video, including transformation, inpainting, extension, upscaling, and reference-based editing. Audio capabilities include local generation, fine-tuning, transcription, text-to-speech, and speech-to-text.