Unsloth

Unsloth is an open-source fine-tuning and RL library built around hand-written Triton kernels that reduce memory use and speed up training on a single GPU. The core Python package is Apache-2.0 (pip version 2026.8.1, released August 2, 2026), with some optional components — notably the new no-code Unsloth Studio UI — under AGPL-3.0. It supports LoRA, QLoRA, full fine-tuning, pretraining, and RL methods such as GRPO across 500+ model families (Llama, Qwen, Gemma, DeepSeek, GLM, Mistral, gpt-oss), and exports to GGUF, 16-bit safetensors, or LoRA adapters.

Its focus is making fine-tuning on your own data viable on hardware most engineers actually have: the project's published floor is any NVIDIA GPU of CUDA capability 7.0+ (T4, RTX 20-series and newer), with QLoRA on a 7B model fitting in about 5 GB of VRAM, and vendor-reported gains of ~2x speed and ~70% less VRAM versus standard Hugging Face + FlashAttention baselines. Dataset and checkpoint handling stay deliberately local — Hugging Face `datasets` plus standard `TrainingArguments` checkpoints on local disk with `resume_from_checkpoint` — with no native S3/object-storage path support documented; the object-storage story here is indirect (users bring their own sync or W&B artifact logging).