Unsloth

Unsloth is an open-source library and local UI for training, fine-tuning, and running open-source language, diffusion, TTS, and embedding models on consumer hardware. Its core claim, backed by published benchmarks, is training up to 2x faster while using 70% less VRAM than standard fine-tuning approaches, with no accuracy loss — achieved through custom Triton kernels and manual backpropagation optimizations rather than approximations that degrade quality. It supports the full range of modern fine-tuning techniques: LoRA, QLoRA, full fine-tuning, pretraining, reinforcement learning (GRPO, DPO), and FP8 training, and works with popular open model families including Qwen, DeepSeek, Gemma, and Llama, plus GGUF and MLX export for local inference. The project is Apache-2.0 licensed and has grown to 67,900+ GitHub stars and 6,100+ forks, making it one of the most widely adopted open-source tools for anyone fine-tuning LLMs without renting large-scale cloud GPU clusters.