unslothai/unsloth

Unsloth is a tool for running and fine-tuning large AI language models on your own computer, with a focus on making this dramatically faster and less demanding on memory. Fine-tuning means taking an already-trained AI model and training it further on your own data so it behaves differently, for example, teaching a general-purpose language model to answer questions in a specific style or domain. The problem Unsloth addresses is that fine-tuning large models typically requires enormous amounts of GPU memory (VRAM) and takes a long time, pricing out anyone without expensive hardware. Unsloth achieves its efficiency gains through custom low-level code optimizations called kernels, which are tuned routines that make the mathematical operations inside neural network training run faster. According to the README it can make training up to 2x faster while using up to 70% less VRAM compared to standard approaches, with no loss in accuracy. It supports over 500 different open-source models including Llama, Gemma, Qwen, DeepSeek, Mistral, and others. There are two ways to use it: Unsloth Studio is a web-based graphical interface you run locally where you can download models, chat with them, and train them through a visual interface, Unsloth Core is the code-based version for more advanced users who want to write training scripts in Python. It supports various training methods including standard fine-tuning, reinforcement learning, and quantized training (reducing model precision to save memory). It runs on NVIDIA GPUs primarily, with macOS and AMD support growing. The tech stack is Python, installable via pip or a one-line shell script.