Unsloth

unsloth.ai

Visit Website

A fine-tuning library that makes QLoRA and LoRA training 2x faster and use 50% less memory than standard implementations, enabling fine-tuning of large models on consumer GPUs.

Why it is useful

Makes fine-tuning models that previously required expensive A100 GPUs practical on a single RTX 3090 or 4090. The speed and memory improvements come from custom CUDA kernels rather than approximations, so accuracy is identical to standard training. For anyone wanting to fine-tune Llama, Mistral, or Qwen on their own data without renting cloud GPUs, this is the starting point.