Unsloth
by Unsloth AI · Framework · also known as unsloth (PyPI), Unsloth Studio, Unsloth Desktop
Open-source framework for fine-tuning and running LLMs with reduced memory, covering LoRA, QLoRA, full fine-tuning and RL, GGUF export, Dynamic quantizations and a local Studio/Desktop UI
Overview
Unsloth is an Apache-2.0 framework for training and running language models on your own hardware with lower memory use than stock implementations, achieved through hand-written kernels and optimized attention paths. It supports LoRA, QLoRA and full fine-tuning, continued pretraining, and reinforcement-learning methods such as GRPO and DPO, extending beyond text to diffusion, text-to-speech and embedding models, with FP8 training on supported GPUs. Trained models export to GGUF, NVFP4, FP8 and other formats for llama.cpp-style runtimes, and the team publishes widely used "Dynamic" GGUF quantizations of popular open models.
In 2026 the project added Unsloth Studio, a browser-based UI, and Unsloth Desktop, a native app for macOS, Windows and Linux installable through Homebrew or WinGet, which run and train models locally and serve them through an OpenAI-compatible API; unsloth start connects coding agents such as Claude Code or Codex to those local models, and Data Recipes build datasets from PDFs, CSVs and DOCX files. Platform support covers NVIDIA, AMD and Intel GPUs, CPU and a Docker image, with free Colab notebooks as the traditional on-ramp. The main site blocks automated clients, and the fast release pace sometimes breaks dependencies between Unsloth, Transformers and PyTorch versions, so pinning matters.
Editorial note on uncertainty. The documentation does not describe paid tiers; pricing_model reflects the open-source software only. unsloth.ai itself could not be loaded by automated clients.
Capabilities
- LoRA, QLoRA, full fine-tuning and continued pretraining
- RL methods including GRPO and DPO; FP8 training
- Diffusion, TTS and embedding model training
- Export to GGUF, NVFP4 and FP8; Dynamic GGUF quantizations
- Unsloth Studio web UI and Desktop app for macOS, Windows and Linux
- OpenAI-compatible local serving and agent connection via `unsloth start`
- Data Recipes for building datasets from documents
Typical workflows
- Fine-tune an open model with QLoRA on a single GPU and export to GGUF for Ollama
- Run GRPO reinforcement learning on a reasoning task
- Serve a local model to Claude Code through the Desktop app
Strengths
- Lower memory and faster training than stock pipelines
- Apache-2.0 license
- Smooth export path to local runtimes
Limitations
- Version coupling with Transformers and PyTorch
- Main website rejects automated clients
- Best performance still requires NVIDIA GPUs
Prerequisites
- Python with PyTorch and a supported GPU, or the Desktop app
- Hugging Face access for base models
Who it is for
- Researchers and students fine-tuning open models on limited hardware
- Developers producing quantized models for local deployment
Alternatives & comparison
- Axolotl
YAML-configured open-source framework for fine-tuning and post-training LLMs: full, LoRA and QLoRA training, preference and RL methods, multimodal models and multi-GPU or multi-node runs
Open sourceOpen source - MLX LM
Python package and CLI for running, serving, quantizing and fine-tuning large language models on Apple Silicon with MLX
Open sourceOpen source - PyTorch
Open-source deep-learning framework with eager tensors, autograd, torch.compile and distributed training, governed by the PyTorch Foundation and the base of most LLM tooling
Open sourceOpen source - Transformers
Hugging Face's model-definition library for text, vision, audio, video and multimodal models, with Pipeline, Trainer and generate APIs used as the reference by training and inference engines
Open sourceOpen source
Sources & verification
Primary sources consulted for this record. Claims are tied to the fields each source supports; the automated link checker re-verifies every URL on a schedule.
- Licensehttps://github.com/unslothai/unsloth/blob/main/LICENSEchecked Nov 1, 2026
Supports: license_spdx, source_model
- Official documentationchecked Nov 1, 2026
- Official siteunslothai/unslothchecked Nov 1, 2026
https://github.com/unslothai/unsloth
Supports: name, summary, capabilities
unsloth.ai root returns 403 to automated clients; the repository is used as homepage
- Confidence
- High
- Reviewed
- Nov 1, 2026
- Links
- Verified · Nov 2, 2026
Spotted an error?
More in Training & fine-tuning
MLX
Apple Inc. · Apple Silicon ML & MLX
Apple's open-source array framework for machine learning on Apple Silicon, with NumPy-like Python, C++ and Swift APIs and unified CPU/GPU memory
Ray
Ray project (Anyscale and contributors) · Distributed & parallel computing
Open-source Python framework for distributing tasks, actors and ML workloads across a cluster, with libraries for data processing, training, tuning, serving and reinforcement learning
Hugging Face Hub
Hugging Face · Training & fine-tuning
Hosting platform for open models, datasets and Spaces demo apps built on git repositories and Xet storage, with Inference Providers, Endpoints, Jobs and an open-source Python client and CLI
scikit-learn
scikit-learn developers · Mathematical & statistical computing
BSD-licensed Python machine-learning library for classification, regression, clustering, dimensionality reduction, preprocessing and model selection on NumPy and SciPy
Sentence Transformers
Hugging Face · Embeddings, reranking & retrieval
Python library for computing and training embedding, cross-encoder reranker, sparse and multi-vector retrieval models, maintained by Hugging Face with thousands of pretrained models on the Hub
Together AI
Together AI · Model providers & APIs
Cloud platform for open-source models: serverless and dedicated inference, fine-tuning, GPU clusters, batch jobs and code sandboxes behind an OpenAI-compatible API