Unsloth Studio
5 hours ago
- #local-models
- #AI-training
- #no-code
- Unsloth Studio is an open-source, no-code web UI for training, running, and exporting AI models locally.
- Supports running GGUF and safetensor models on Mac, Windows, and Linux with 2x faster training and 70% less VRAM.
- Auto-creates datasets from PDF, CSV, JSON, DOCX, and TXT files without needing a pre-existing dataset.
- Features include self-healing tool calling, web search, code execution, and auto inference parameter tuning.
- Supports text, vision, TTS audio, and embedding models with multi-GPU inference and most model types.
- No-code training allows uploading documents or YAML configs to start training instantly on NVIDIA GPUs.
- Data Recipes transform unstructured files into usable datasets via graph-node workflow.
- Observability features track training loss, gradient norms, and GPU utilization in real-time.
- Export models to safetensors or GGUF for use with llama.cpp, vLLM, Ollama, and more.
- Model Arena allows comparing outputs of different models side by side.
- Beta version currently available with improvements and new features coming soon.
- Works on Windows, Linux, WSL, and MacOS (chat only for Mac currently).
- Training requires NVIDIA GPUs; CPU-only supports chat inference.
- Future support planned for Apple MLX, AMD, and Intel.
- Includes a free Google Colab notebook for exploring features on T4 GPUs.
- Workflow includes loading models, importing data, refining datasets, training, and exporting models.
- Unsloth Studio runs 100% offline and locally, with no data collection.
- Dual-licensed under Apache 2.0 and AGPL-3.0 for different components.
- Supports a range of model families beyond LLMs, including multimodal and audio models.
- Future plans include multi-GPU support and collaborations with NVIDIA.