Unsloth
An open-source local LLM app for macOS, Windows and Linux. Run and train models, generate media, and connect coding agents to your hardware.
Unsloth brings model training and everyday AI use into a desktop app for people who want to run models on their own hardware. Its no-code interface covers chat, fine-tuning and media generation on macOS, Windows and Linux. The Unsloth software is open source under Apache 2.0.
Model support includes GGUF and MLX, with text models such as Qwen3.8, DeepSeek-V4 and Gemma 4 alongside diffusion, audio and embedding models. A built-in model hub helps you find and download a quantization suited to your device. It supports CPUs, NVIDIA, AMD and Intel GPUs, plus multi-GPU setups and Vulkan.
Training goes beyond adapting a chat model. Unsloth supports LoRA, QLoRA, full fine-tuning and reinforcement learning, and can build datasets from PDFs, CSVs and DOCX files. You can export trained models in formats including GGUF. For media work, it supports image generation and editing with Qwen-Image-2.1 and FLUX, plus video generation with Wan and LTX.
Developers can connect Claude Code, Codex and MCP to local models, or use an OpenAI-compatible API with existing apps. Models can execute Bash and Python in a sandbox, with automatic repair and retries for failed tool calls.
Desktop inference, training and media generation run locally. Web search and research with cited reports use online access. Unsloth also offers a self-hosted web UI through Docker and a code-based edition; optional Cloudflare tunnels provide remote access to models running locally or on Colab.
