Unsloth runs and fine-tunes open models on local hardware. It covers LLMs, MLX and GGUF models, diffusion, embedding, text-to-speech, and audio models, and it runs on Windows, Linux, WSL, and macOS with NVIDIA, AMD, and Intel GPUs, multi-GPU setups, CPUs, and a Vulkan backend.
It comes in three forms: Unsloth Desktop (a native app), Unsloth Studio (a web UI), and Unsloth Core (the Python library for code-based training). Free Colab notebooks cover fine-tuning and reinforcement learning for models such as Gemma 4, Qwen3.5, and gpt-oss.
For agent builders, Unsloth serves local models through OpenAI- and Anthropic-compatible APIs, and Unsloth Start connects coding agents to those models with one command. Supported agents include Claude Code, Codex, DeepSeek Harness, Hermes, Mistral Vibe, OpenCode, and OpenClaw. Local models also connect to MCP servers for tool calling and code execution, and the Connections feature mixes local models with API providers or servers such as vLLM and Ollama in the same interface.
Training uses custom Triton kernels and runs model training up to 2× faster with 70% less VRAM, with mixture-of-experts training up to 12× faster. Models run and save through the llama.cpp library, and trained models export to GGUF, NVFP4, FP8, and other formats.
Features
- Fine-tuning methods: LoRA, QLoRA, full fine-tuning, pretraining, and FP8 training
- Reinforcement learning: GRPO, GSPO, DPO, and vision RL, including long-context RL
- Model coverage: Qwen, GLM, Kimi, DeepSeek, Gemma, MiniMax, and gpt-oss families, plus image and video diffusion models
- Unsloth Start: points Claude Code, Codex, Hermes, OpenCode, OpenClaw, and other agents at a local model
- Compatible APIs: serves local models over OpenAI- and Anthropic-compatible endpoints
- MCP and tools: local models call MCP servers, run code, and use web search, deep research, and RAG
- Remote access: reach local models from other devices on the LAN or through a Cloudflare HTTPS link
- Export: save trained models as GGUF, NVFP4, FP8, and other formats
- Data Recipes: build training datasets from PDFs, CSVs, DOCX files, and other documents
- Hardware support: NVIDIA, AMD, and Intel GPUs, multi-GPU setups, CPUs, and Vulkan
