Unsloth

Fine-tuning open-weight models on a single workstation is the main job here, and the project now ships in three forms: Unsloth Core, a Python package that replaces stock training layers with custom Triton kernels; Unsloth Studio, a browser UI started from the command line; and Unsloth Desktop, a packaged app for Windows, macOS and Linux. Supported methods include LoRA and QLoRA adapters, full-parameter training, continued pretraining and reinforcement learning recipes such as GRPO and DPO, across text LLMs plus image diffusion, speech synthesis and embedding models. The maintainers claim about twice the speed of standard pipelines with far lower VRAM use. Weights export to GGUF, FP8 or NVFP4, and the app can also chat with local models, expose an OpenAI compatible endpoint and launch coding agents like Claude Code or Codex against a local model. NVIDIA cards are the primary target, with AMD, Intel, CPU and Vulkan routes documented; install through a script, Homebrew, Docker or uv. Licensing is split: the core library is Apache-2.0, while Studio code is AGPL-3.0-only, which matters if you plan to host or embed the UI.