Unsloth is an open-source framework designed to accelerate the training and inference of Large Language Models (LLMs), diffusion, embedding, and audio models. The platform provides three deployment modes: Unsloth Desktop, Unsloth Studio (a web UI), and Unsloth Core for code-based workflows. Technically, the engine optimizes fine-tuning through methods like LoRA, QLoRA, and full fine-tuning, supporting reinforcement learning techniques such as GRPO and DPO. Performance benchmarks indicate significant efficiency gains: training can be 2× faster with 70% less VRAM for models like Llama 3.1 (8B), while MoE LLMs can see 12× faster training with 35% less VRAM. The software supports a wide array of hardware, including NVIDIA, AMD, and Intel GPUs, as well as CPUs via the Vulkan backend. Users can run diverse models such as Qwen3.8, Gemma 4, and DeepSeek-V4, with the ability to export to formats like GGUF and FP8. Additionally, Unsloth enables local model integration with agents like Claude Code and OpenAI Codex through OpenAI-compatible APIs and supports RAG and private web search.