Unsloth

Unsloth provides an open-source framework and desktop application designed to accelerate local large language model fine-tuning and inference. Operating entirely offline on macOS and Windows, the interface allows developers to execute GGUF and Safetensors models with integrated tool-calling, web browsing, and OpenAI-compatible API endpoints. The system includes custom Triton kernels that optimize memory usage and training speed for techniques such as LoRA, QLoRA, and full fine-tuning across hundreds of vision, text, and multimodal architectures. Users can automatically generate structured datasets from raw documents, CSV files, and PDFs using visual graph workflows. The environment supports local execution of Python and Bash scripts, allowing models to run code natively during evaluation. Fine-tuned weights can be exported directly into popular formats like GGUF or Safetensors for immediate deployment across edge frameworks. While the local setup eliminates cloud dependence and drastically reduces memory overhead, current releases do not support multi-node distributed training, restricting multi-GPU scaling for trillion-parameter foundation models.