Unsloth

Unsloth is an open-source framework for fine-tuning and running open models, now also shipped as a desktop application so that neither step requires assembling a Python environment. Its emphasis is memory efficiency — the claim is faster training with less VRAM, which is what decides whether a fine-tune fits on the hardware you already own — and it covers text, image, text-to-speech and embedding models, with reinforcement learning as well as supervised fine-tuning.

The reason memory efficiency is the whole product here is arithmetic. Fine-tuning an open model does not only require the weights; optimiser state and activations push the requirement well past the model file, so a model that runs comfortably for inference can be out of reach for training on the same card. Techniques that cut that overhead change what is possible rather than merely what is convenient, and the honest caveat is that they trade against something — attention to numerics, or compatibility with unusual architectures — so the speedup is real but not free. The desktop app is new and the framework's supported-model list is the practical limit on what you can do.