Unsloth

Unsloth is an open-source platform with 78,000+ stars on GitHub for running and fine-tuning LLMs locally — famous for making model training dramatically faster and lighter on consumer GPUs.

Overview

Unsloth started as the fine-tuning library that let a single consumer GPU train models that normally demanded a datacenter card, using hand-optimized kernels to cut memory use and speed up training for Llama, Qwen, Gemma, DeepSeek and other open models. It has since grown into a full local AI studio: a desktop UI to download, run and train LLMs and diffusion models, with support for GGUF and MLX formats, reinforcement learning, text-to-speech and image models like FLUX. Newer additions include an OpenAI-compatible API, the ability to connect coding agents like Claude Code to local models with a single command, sandboxed code execution and built-in web search and deep research. Everything runs on your own hardware, and the project — Apache-2.0 licensed — is updated constantly, with repository activity in October 2026 and day-zero support for new model releases.

Key Features

• Efficient fine-tuning – train 100+ model families with far less VRAM and much faster throughput.
• Local studio UI – download, chat with and train models from one desktop interface.
• Broad format support – GGUF and MLX, plus QLoRA, LoRA and full fine-tuning workflows.
• Agent connections – point Claude Code and other agents at your local models.
• OpenAI-compatible API – serve local models to existing apps and scripts.
• More than text – diffusion image models, TTS and multimodal models in the same studio.

💳 Pricing

Unsloth is free and open source under the Apache License 2.0. Running and training locally costs nothing beyond your own hardware and electricity.

Who Is It For

Developers, researchers and hobbyists who fine-tune open models on limited hardware, and anyone who wants a friendly local studio for running models privately.