Unsloth Studio is a beta-stage, open-source web interface designed to simplify local AI model management. It enables users to run GGUF and safetensor models on Mac, Windows, and Linux, and train over 500 models 2x faster with 70% less VRAM using optimized kernels for LoRA, FP8, FFT, and PT. Key features include:
- Local Model Execution: Search and run models with self-healing tool calling, web search, auto inference parameter tuning, and code execution via llama.cpp and Hugging Face integrations.
- No-Code Training: Upload PDF, CSV, JSON, or YAML files to start training instantly on NVIDIA GPUs, with support for multi-GPU setups and models like Qwen3.5 and NVIDIA Nemotron 3.
- Data Recipes: Transform unstructured documents into synthetic datasets using graph-node workflows powered by NVIDIA DataDesigner.
- Observability: Real-time tracking of training loss, gradient norms, and GPU utilization with customizable dashboards accessible on multiple devices.
- Model Export: Save or export models to safetensors or GGUF formats for use with llama.cpp, vLLM, Ollama, and LM Studio.
- Model Arena: Compare two models side-by-side to evaluate differences in outputs.
- Privacy-First Design: Operates 100% offline with token-based authentication (password and JWT) for secure, local data control.
Target users include developers, researchers, and businesses seeking efficient, private AI model fine-tuning and deployment without cloud dependencies. The tool supports text, vision, TTS audio, and embedding models, with future plans for Apple MLX, AMD, and Intel training support.
