unslothai/unsloth is a tool for running and training LLM and diffusion models locally. The repository description states that it supports the GGUF and MLX formats as well as Qwen3.8, DeepSeek-V4, MiniMax-H3, Gemma 4, and FLUX. The most recent release is v0.1.909-beta.
The README describes this repository as a desktop app for running and training models. It comes in three forms: the Unsloth Desktop desktop app, Unsloth Studio, a browser-based web UI, and the Unsloth Core version for writing code directly. It supports multi-GPU configurations and NVIDIA, AMD, and Intel GPUs, CPUs, and the Vulkan backend on Windows and Linux, WSL, and macOS. The models it covers are broad, spanning LLMs as well as diffusion, decision, embedding, and audio models, and it supports LoRA, QLoRA, and reinforcement learning alongside fine-tuning, leading on to export and deployment. LoRA is a method that keeps the base model frozen and trains only additional weights to adapt it. The code is written in Python, the license is Apache-2.0, and the repository has 78,322 GitHub stars. It was created on November 29, 2023.
However, the research record lists only a momentum score of 7.0 on a 0-10 scale, and no novelty, technology, or quality scores are recorded. The available information contains no figures on execution speed or training time, and no comparison results against other tools either. Nor can it be confirmed what v0.1.909-beta changed. Rather than adopting it right away while the changes are unknown, it is better to check the release notes first.