🦥Unsloth Docs

Unsloth lets you run and train AI models on your own local hardware.

Our docs will guide you through running & training your own model locally.

Get started Our GitHub

Introducing Unsloth Studio

A new open source, no-code web UI to train and run LLMs.

Run models with Unsloth Studio

Run GGUF models on your local Mac, Windows and Linux.

Qwen3.5

New Qwen3.5 Small & Medium LLMs are here!

NVIDIA Nemotron 3

Run the new 4B and 120B models by NVIDIA.

Faster MoE is here!

Train MoE LLMs 12x faster with less VRAM.

Claude Code & Codex

Learn to run local LLMs via Claude & OpenAI.

Qwen3-Coder-Next

Run & fine-tune the new 80B coding model.

GLM-4.7-Flash

Run & fine-tune 30B model for agentic coding.

🦥 Why Unsloth?

• We directly collab with teams behind gpt-oss, Qwen3, Llama 4, Mistral, Gemma 1-3 and Phi-4, where we’ve fixed critical bugs that greatly improved model accuracy.

• Unsloth streamlines local training, inference, data, and deployment

• Unsloth supports inference and training for 500+ models: vision, TTS, embedding, RL

⭐ Features

Unsloth lets you run and train models for text, audio, embedding, vision and more. Unsloth provides many key features for both inference and training:

Inference

• Search + download + run any model like GGUFs, LoRA adapters, safetensors.

• Self-healing tool calling / web search and call OpenAI-compatible APIs.

• Auto inference parameter tuning and edit chat templates.

• Export or save your model to GGUF, 16-bit safetensor etc.

• Compare outputs with two different model side by side.

Training

• Train 500+ models ~2x faster with ~70% less VRAM (no accuracy loss)

• Supports full fine-tuning, pre-training, 4-bit, 16-bit and FP8 training.

• Auto-create datasets from PDF, CSV, DOCX files. Edit data in a visual node workflow.

• Observability: Monitor training live, track loss, GPU usage, customize graphs

• Most efficient reinforcement learning library, using 80% less VRAM for GRPO, FP8 etc.

• Multi-GPU works but a much better version is coming!

Quickstart

Unsloth supports MacOS, Linux, Windows, NVIDIA and CPU setups. See: Unsloth Requirements

MacOS, Linux, WSL:

Windows PowerShell:

Docker

Use our official Docker image: unsloth/unsloth which currently works for Windows, WSL and Linux. MacOS support coming soon.

Launch Unsloth

New Models

Kimi K2.5

MiniMax-M2.5

GLM-5

What is Fine-tuning and RL? Why?

Fine-tuning an LLM customizes its behavior, enhances domain knowledge, and optimizes performance for specific tasks. By fine-tuning a pre-trained model (e.g. Llama-3.1-8B) on a dataset, you can:

• Update Knowledge: Introduce new domain-specific information.

• Customize Behavior: Adjust the model’s tone, personality, or response style.

• Optimize for Tasks: Improve accuracy and relevance for specific use cases.

Reinforcement Learning (RL) is where an "agent" learns to make decisions by interacting with an environment and receiving feedback in the form of rewards or penalties.

• Action: What the model generates (e.g. a sentence).

• Reward: A signal indicating how good or bad the model's action was (e.g. did the response follow instructions? was it helpful?).

• Environment: The scenario or task the model is working on (e.g. answering a user’s question).

Example fine-tuning or RL use-cases:

• Enables LLMs to predict if a headline impacts a company positively or negatively.

• Can use historical customer interactions for more accurate and custom responses.

• Fine-tune LLM on legal texts for contract analysis, case law research, and compliance.

You can think of a fine-tuned model as a specialized agent designed to do specific tasks more effectively and efficiently. Fine-tuning can replicate all of RAG's capabilities, but not vice versa.