Unsloth
Open-source platform for fine-tuning and running LLMs locally — 2x faster training with 70% less VRAM. Supports LoRA, QLoRA, DPO, GRPO.
Unsloth is an open-source platform for fine-tuning and running large language models locally. Built on PyTorch with custom Triton and CUDA kernels, it claims up to 2x faster training while using 70% less VRAM than standard Hugging Face + TRL setups. The project supports the full fine-tuning workflow — LoRA, QLoRA, full fine-tuning, continued pretraining, DPO, GRPO, and reinforcement learning — across transformer models from Llama, Mistral, Qwen, Gemma, DeepSeek, and others.
The Unsloth Desktop app wraps the core library in a cross-platform GUI (Windows, macOS, Linux) with one-click model downloads, an OpenAI-compatible local API server, and tools for building datasets from PDFs, CSVs, and other documents. Unsloth Studio is the more featured web UI, with support for AMD GPUs, MLX on Apple Silicon, and a chat interface. The project also ships Unsloth Start, which connects local models to coding agents like Claude Code, Codex, and OpenCode as drop-in subagents.
Unsloth is dual-licensed: the core Python library and CLI ship under Apache 2.0, while the Unsloth Studio UI is AGPL-3.0. Models trained with Unsloth can be exported to GGUF, NVFP4, FP8, and standard PyTorch formats for deployment in llama.cpp, vLLM, Ollama, or back into Hugging Face Transformers. The project also publishes pre-quantized Dynamic GGUF models on Hugging Face.
