Matt Mireles released a toolkit enabling fine-tuning of Gemma 4 and 3n models for text, image, and audio modalities directly on Apple Silicon Macs. This addresses a critical gap in local AI development, as traditional methods require expensive NVIDIA GPUs and significant local storage. The system streams training data from cloud storage like Google Cloud Storage (GCS) or BigQuery, allowing users to leverage terabytes of data without filling their SSD. Gemma 4 and Gemma 3n support LoRA fine-tuning for tasks like captioning, VQA, and domain-specific ASR, running entirely on Apple's Metal Performance Shaders (MPS) for native performance.

Mireles highlights the challenge of memory limits on his 64GB Mac Studio, noting frequent OOM errors during longer sequence training. The toolkit provides a wizard interface and detailed configuration for setting up text-only, image-text, or audio-text fine-tuning profiles.