Unsloth
Run and fine-tune open models from a desktop app, a web UI or code, with GGUF export and day-one support for many new releases.
Unsloth is a project for running and training open models. It can be used in three ways: Unsloth Desktop, a desktop app; Unsloth Studio, a web interface; and Unsloth Core, the code-based package for scripts and notebooks. It covers language, MLX, GGUF, diffusion, embedding and audio models, and its documentation tracks new releases closely, including Qwen3.8, GLM-5.3-Flash, Kimi K3, Qwen-Image-2.1, DeepSeek V4 and Gemma 4.
For training, you can fine-tune a model on your own data, build datasets from PDFs, CSVs and Word files with its Data Recipes feature, and then export the result as GGUF, NVFP4, FP8 and other formats for deployment. It runs on Windows, Linux, WSL and macOS, with support for multi-GPU setups, NVIDIA, AMD and Intel GPUs, CPUs and Vulkan.
Unsloth uses dual licensing. The core package is Apache 2.0, while optional components such as the Studio UI are AGPL-3.0, which asks you to share the source of modified versions you offer to users over a network. Check which part you use before building a product around it.
Good for
- Fine-tuning a small or medium open model on your own data without a cluster
- Preparing a model for local use, for example by exporting GGUF files
- People who prefer a graphical interface for training experiments
Watch out for
- The Apache 2.0 and AGPL-3.0 split matters if you offer the Studio UI as a service
- Fine-tuning still needs enough GPU memory for the model you choose
- The project changes quickly, and the releases are still tagged beta