unslothai/unsloth

What it solves

Unsloth is a comprehensive toolset for running and training AI models locally on a variety of hardware (NVIDIA, AMD, Intel GPUs, CPUs, and Vulkan). It addresses the high VRAM requirements and slow training speeds typically associated with fine-tuning large models, allowing users to train models up to 2x faster with 70% less VRAM without losing accuracy.

How it works

Unsloth provides three primary interfaces: a native desktop application (Unsloth Desktop), a web-based UI (Unsloth Studio), and a code-based library (Unsloth Core). It supports a wide range of model types, including LLMs, diffusion, embedding, and audio models. For training, it utilizes optimized kernels and techniques like LoRA, QLoRA, and GRPO to reduce resource consumption. It also includes "Data Recipes" to help users build datasets from common file formats like PDFs and CSVs.

Who it’s for

It is designed for developers and AI researchers who want to run and fine-tune frontier models locally on their own hardware, as well as those who want to connect local models to agentic frameworks like Claude Code or Codex.

Highlights

• Hardware Versatility: Supports Windows, Linux, macOS, and WSL across NVIDIA, AMD, and Intel GPUs.
• Efficient Training: Offers 2x faster fine-tuning with significantly reduced VRAM usage.
• Broad Model Support: Compatible with LLMs (e.g., Qwen, DeepSeek, Gemma), diffusion models, and TTS.
• Agent Integration: The unsloth start command allows local models to be used with agents like Claude Code and Codex via OpenAI-compatible APIs.
• Deployment Flexibility: Supports exporting models to formats like GGUF, NVFP4, and FP8.
• RAG & Search: Includes built-in private web search, deep research, and RAG capabilities.