Features
No-code fine-tuning of LLMs and SLMs
Support for H2O Danube3 and H2OVL Mississippi models
Automatic hyperparameter optimization
Built-in evaluation metrics and leaderboards
Import datasets in CSV, JSON, Parquet, and other formats
Export fine-tuned models for deployment
Integration with H2O MLOps for model lifecycle management
On-premise and air-gapped deployment
Multi-model support with cost controls
Preconfigured training recipes for common tasks
Visual interface for training configuration
Human-in-the-loop evaluation support
Custom CUDA kernels for LoRA, FP8, full fine-tuning
Up to 2x faster training vs Flash Attention 2
Up to 90% less memory usage vs Flash Attention 2
Unsloth Studio no-code visual interface (offline on Mac/Windows)
Data Recipes auto-create datasets from PDF, CSV, JSON
Model Arena side-by-side model comparison
Export to GGUF and safetensors
OpenAI-compatible API endpoint
Tool-calling and web search capability
Run 100% offline on Mac and Windows
Supports 500+ models (text, vision, audio, embeddings)
Reinforcement learning (GRPO) with 80% less VRAM
Multi-GPU support (Pro, coming to Free)
Multi-node support (Enterprise only)
Quantization-Aware Training (QAT)
