Gemma fine‑tuning on Apple Silicon

Training a model usually means changing millions or billions of internal weights, like retuning every knob in a giant soundboard. Fine-tuning is the cheaper version: you keep the base model and adjust a much smaller layer so it learns your niche task without starting from zero. (ai.google.dev) Google’s Gemma family is built for that kind of adjustment. Google said on April 2, 2026 that Gemma 4 was sized to run and fine-tune on hardware ranging from Android devices to laptops and workstations, not just data-center servers. (blog.google) Apple Silicon changes the hardware side of the equation. Apple’s M-series chips use unified memory, which means the central processor and graphics processor share one pool of memory instead of copying data back and forth like two cooks passing one cutting board across a kitchen. (opensource.apple.com) That shared-memory design is why Macs have become usable for local machine learning. Apple’s Metal Performance Shaders backend lets PyTorch run training workloads on Mac graphics processors, so model work that once needed a separate NVIDIA setup can run on a laptop. (developer.apple.com) The new piece is multimodal fine-tuning on that hardware. A public Gemma Multimodal Fine-Tuner repository published this week says Gemma 4 and Gemma 3n can now be fine-tuned on Apple Silicon for text, image-plus-text, and audio-plus-text tasks using PyTorch and Metal Performance Shaders. (github.com) It uses Low-Rank Adaptation, usually shortened to LoRA after you have heard the long name once, which trains a small set of extra weights instead of rewriting the whole model. The repository says that approach works for captioning, visual question answering, instruction tuning, and audio-text training on a Mac. (github.com) The audio part is the detail that stands out. The project’s comparison table says audio-plus-text fine-tuning works on Apple Silicon here, while several common alternatives either do not support audio at all or depend on NVIDIA’s CUDA software stack. (github.com) There is also a storage angle. The same tool says it can stream training data from Google Cloud Storage and BigQuery, so a developer can train on datasets larger than the solid-state drive inside a MacBook instead of downloading everything first. (github.com) That fits the direction Google has been pushing Gemma. Its official fine-tuning docs already point developers to parameter-efficient tuning methods and say some behavior changes can show up with as few as 20 prompt-response pairs, which is closer to prototyping than to building a giant foundation model from scratch. (ai.google.dev) Apple has been building the software rails for this from the other side. Its MLX framework is described by Apple as a machine learning array framework for Apple Silicon, optimized for unified memory, which is exactly the kind of setup local model developers want when they are testing ideas on one machine. (github.com) So the practical shift is not that Macs suddenly beat server clusters. It is that a developer with an M-series laptop can now try a Gemma text model, an image model, or an audio model locally, change it with LoRA, and see whether the idea works before paying for a cloud training run. (blog.google) (github.com)