Gemma Tuner Multimodal

Fine-tune Gemma 3n and 4 models on your Mac using text, images, and audio without needing an NVIDIA GPU. This runs entirely on Apple Silicon through Metal Performance Shaders and supports LoRA parameter-efficient training. The standout feature is streaming datasets from GCS or BigQuery so you can train on huge datasets without copying everything locally. You get a CLI with a wizard for setup, separate modes for instruction tuning versus completion, and proper multimodal support including audio transcription, which is rare for local Mac training. If you have domain-specific data and an M-series Mac, this gives you a complete fine-tuning pipeline that actually runs locally instead of just theoretically supporting it.