Run and train the latest AI models locally without the cloud
You can run models like Qwen3.8 or DeepSeek-V4 directly on your computer without relying on external servers, and train them if needed. Unsloth is a desktop app that bundles tools for model execution and training. You can operate it through the interface without writing code, or use code for more precise control.
Connecting local models and agents
Unsloth is a native desktop app that runs on Windows, macOS, and Linux. Once installed, you can handle model downloading, execution, and training from a single screen. The 'Unsloth Start' feature lets you connect agent tools like Claude Code or Codex directly to local models. For example, the unsloth start claude command configures Claude Code to use the Qwen3.8 model on your computer. Since all processing happens locally without external API calls, it is suitable for environments that prioritize data privacy.
Supported models and training methods
It supports not only text generation models (LLMs) but also image generation, speech synthesis, and embedding models. It handles various models such as Qwen3.8, Gemma 4, DeepSeek-V4, and FLUX, and supports multiple formats including GGUF, MLX, NVFP4, and FP8. Training features include latest techniques such as LoRA, QLoRA, full fine-tuning, reinforcement learning, GRPO, and DPO. According to the README, this method trains models with no loss in accuracy while being 2x faster and reducing VRAM usage by up to 70%. It also includes a 'Data Recipes' feature that converts PDF or CSV files into datasets, simplifying the data preparation process.
Installation and system requirements
The easiest way is to download the desktop app for your operating system. Versions are available for Windows, macOS, and Linux (Deb, AppImage, ARM64). If you prefer working with code, you can install 'Unsloth Core', which requires a Python 3.13 environment. On Linux and WSL, you can install it using the curl command, while on Windows, you can use a PowerShell script. GPU support covers NVIDIA, AMD, and Intel GPUs, as well as CPUs and the Vulkan backend. It can also run in container environments via Docker images; if you use an AMD GPU, you must use the unsloth/unsloth-rocm image. When exposing it to the network, you can generate a Cloudflare HTTPS link using the --secure option or configure LAN access.
License and security checks
Unsloth applies a dual license of Apache 2.0 and AGPL-3.0. The core package is under Apache 2.0, but the web UI, 'Unsloth Studio', follows the AGPL-3.0 license. This means you must verify license conditions for commercial use. Additionally, server-side tools are enabled by default, so when exposing it to the network, you should set an administrator password or use the --disable-tools option to enhance security. You can try training models on free notebooks on platforms like Kaggle, so it is recommended to test this feature before installing on actual hardware.
