Unsloth

Unsloth makes fine-tuning large language models practical. Where HuggingFace trainers need massive GPU clusters and hours of waiting, Unsloth gets the same results in half the time on half the hardware. The library works by rewriting the training loop from scratch. It replaces PyTorch's autograd with hand-derived math for Llama, Mistral, Qwen, Gemma, and other popular architectures. These manual derivatives skip unnecessary computation, which is where the 2x speedup comes from. The memory savings come from custom CUDA kernels that handle 4-bit quantization, gradient checkpointing, and optimizer states more efficiently than generic PyTorch implementations. Setting up Unsloth takes two lines in a Colab notebook. It integrates directly with HuggingFace's ecosystem — you load models the same way, use the same dataset format, and export to the same formats. The difference is speed and memory. A Llama 3.1 8B model that needs 24GB VRAM with standard trainers runs in under 7GB with Unsloth. That means you can fine-tune on a single T4 GPU instead of an A100. Unsloth supports LoRA, QLoRA, and full fine-tuning. LoRA is the default and works well for most use cases. QLoRA with 4-bit quantization pushes memory requirements even lower. Full fine-tuning is available for when you need to modify every parameter. The library also handles DPO, ORPO, and RLHF training, not just supervised fine-tuning. This makes it a single tool for the entire alignment pipeline. You can go from a base model to a chat-tuned, preference-aligned model without switching frameworks. Unsloth is free and open source under Apache 2.0. The Pro tier ($9.99/month) adds faster kernels, longer context support, and priority support. The team releases updates frequently, adding support for new models within days of their release.