What Unsloth does
Unsloth is an open-source library for fine-tuning, reinforcement learning, and running large language models efficiently. Through custom GPU kernels and optimized LoRA/QLoRA implementations, it claims training that is multiple times faster while using up to 70-90% less memory, letting developers fine-tune models on a single consumer GPU. It supports 500+ models including Llama, Mistral, Gemma, and Qwen, and provides quantized GGUF and Safetensors exports plus an OpenAI-compatible API for running models locally.
Adoption and backing
Founded in 2023 by Australian brothers Daniel and Michael Han, Unsloth went through Y Combinator's Summer 2024 batch and is backed by GitHub Accelerator and Microsoft's M12 fund among others. The project has earned tens of thousands of GitHub stars and reports over 10 million monthly model downloads. Unsloth partners with NVIDIA, Google, Meta, Mistral, and others on bug fixes and quantization, and offers a free open-source tier alongside Pro and Enterprise tiers with multi-GPU and multi-node training and faster inference.