Unsloth Documentation & Notebooks
by Unsloth
About 150 press-play notebooks that fine-tune modern open models on a free Colab GPU.
Overview
Unsloth's documentation is the practical manual for its fine-tuning stack, which reaches its speed and memory gains through hand-written Triton kernels and manually derived backward passes rather than autograd-generated ones, combined with gradient checkpointing and Cut Cross Entropy. The docs cover getting started with training and running open models locally, then hand off to a library of roughly 150 ready-to-run notebooks grouped as Colab supervised fine-tuning, GRPO reasoning RL, text-to-speech, speech-to-text, vision and multimodal, embedding models, large LLMs beyond free-Colab VRAM, Kaggle variants and specific use cases, spanning the Gemma, Qwen, Llama, Mistral, DeepSeek and Phi families from 270M to 120B+ parameters. Independent benchmarking published in July 2026 puts the gains at roughly 2x on Llama 3.1 8B and 70B on a single GPU, 6 GB of VRAM for an 8B QLoRA run and 41 GB for 70B, with its sharpest advantage in context headroom: 78,475 tokens for Llama 3.1 8B on a 24 GB card versus 5,789 with stock Transformers plus FlashAttention-2. The same review names the limitation — multi-GPU works through Accelerate and DeepSpeed, but Unsloth does not offer the composable tensor, context and expert parallelism that Axolotl and TRL document, and the kernel gains only apply to architectures explicitly supported. The core package is Apache-2.0 with roughly 69,600 GitHub stars and active daily development, and it now also ships Unsloth Studio, an AGPL-3.0 no-code local web UI for training, running and exporting models on NVIDIA, Intel, Apple and AMD hardware.
At a Glance
- Topic
- Fine-Tuning
- Level
- Intermediate
- Format
- Documentation
- Cost
- Free
- Duration
- Self-paced; most notebooks run end to end in a single Colab session
- Provider
- Unsloth
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Fine-tune an open model with LoRA or QLoRA inside a single Colab session
- ✓Choose a base model and dataset format that suit your VRAM budget
- ✓Run GRPO reinforcement learning to train reasoning behaviour into a model
- ✓Export a tuned adapter to GGUF or Ollama for local inference
- ✓Fine-tune vision, speech-to-text, text-to-speech and embedding models from notebooks
- ✓Understand where the kernel-level optimizations apply and where they silently do not
- ✓Use Unsloth Studio's no-code local UI to train, run and export models
Highlights
- •Roughly 150 ready-to-run notebooks — pick your model and press play
- •Independent July 2026 benchmarking measured ~2x speedups and 6 GB for an 8B QLoRA run
- •Context headroom is the standout: 78,475 tokens on 24 GB versus 5,789 with stock Transformers
- •Apache-2.0 core with roughly 69,600 GitHub stars and daily commits
- •Honest caveat: no composable tensor, context or expert parallelism, and unsupported architectures get no kernel gains
Who It’s For
Best For
- ✓Practitioners fine-tuning open models on a single consumer or rented GPU
- ✓Teams shipping a tuned model to local inference via GGUF or Ollama
- ✓Engineers experimenting with GRPO reasoning training without a cluster
- ✓Anyone who wants a working fine-tune before learning the theory behind it
Prerequisites
- •Basic Python and comfort running Jupyter or Colab notebooks
- •A rough sense of what LoRA and quantization do
- •Access to a GPU — a free Colab session is enough for most notebooks
FAQ
What is Unsloth Documentation & Notebooks?
Unsloth's docs and notebook library for fast, low-memory fine-tuning of open models. Roughly 150 ready-to-run notebooks cover LoRA and QLoRA, GRPO reasoning RL, vision, speech and embedding models across the Gemma, Qwen, Llama, Mistral and DeepSeek families — most small enough to train on a free Colab GPU and export straight to GGUF or Ollama for local inference.
Is Unsloth Documentation & Notebooks free?
Unsloth Documentation & Notebooks is free to access.
What level is Unsloth Documentation & Notebooks for?
Unsloth Documentation & Notebooks is aimed at a intermediate audience. Recommended background: Basic Python and comfort running Jupyter or Colab notebooks, A rough sense of what LoRA and quantization do, Access to a GPU — a free Colab session is enough for most notebooks.
How long does Unsloth Documentation & Notebooks take?
Expect roughly Self-paced; most notebooks run end to end in a single Colab session. Most learners work through it at their own pace.
What will I learn from Unsloth Documentation & Notebooks?
You'll learn: Fine-tune an open model with LoRA or QLoRA inside a single Colab session; Choose a base model and dataset format that suit your VRAM budget; Run GRPO reinforcement learning to train reasoning behaviour into a model; Export a tuned adapter to GGUF or Ollama for local inference; Fine-tune vision, speech-to-text, text-to-speech and embedding models from notebooks; Understand where the kernel-level optimizations apply and where they silently do not; Use Unsloth Studio's no-code local UI to train, run and export models.
Topics
Sources
This page was written from 4 sources, 2 on domains other than unsloth.ai.