Unsloth

Unsloth is an open-source project and startup building a library for fast, memory-efficient fine-tuning and reinforcement learning (RL) of large language models. The tool relies on custom, optimized kernels that, according to its creators, make model training several times faster while requiring substantially less GPU memory. The library supports popular model families including Llama, Gemma, Mistral, Qwen, Phi and DeepSeek, as well as methods such as LoRA and QLoRA. The project is released under the Apache 2.0 license and published openly on GitHub, where it has accumulated tens of thousands of stars. The company was founded in 2023 by brothers Daniel Han and Michael Han. Unsloth went through the Y Combinator accelerator as part of the Summer 2024 batch. The organization is headquartered in San Francisco, California. Beyond the open-source library, the project develops further developer tools centered on training and running AI models locally.