You can turn a 270 million-parameter Gemma 3 model into a specialized code reviewer on any Apple Silicon Mac in under ten minutes. Gemma 3 270M, released by Google in August 2025, loads in about 830 MB of RAM and pushes out 150+ tokens per second on an M3. The trick is MLX’s “uv” toolchain: a single pip install and a couple of shell commands give you a reproducible Python 3.11 environment, automatic model download, and a hassle-free fine-tuning script. The heart of the demo is a tiny dataset of roughly sixty Python code snippets with expert review texts. You write a prepare_data.py that pairs bad code with “ISSUES FOUND” bullet lists—anything from SQL injections in sqlite3 queries to unclosed files and insecure shell commands. Then you fire off uv commands to load the Gemma 3 weights from Hugging Face and train for a handful of epochs. On an M1 Air the process wraps in three minutes; on a larger dataset it might stretch near ten. Once fine-tuned, the model picks up security holes, performance snags, missing imports and stylistic fixes just like a junior engineer would—but entirely offline, with no GPU rental fees. It’s not a replacement for real code audits, but it shows how accessible LLM customization has become: small models, few samples, a pocket laptop, and you’re in business.