The author trained a 67-million-parameter transformer model on an M4 Mac Mini without a discrete GPU, achieving a 93.94% exact-match accuracy in generating CLI commands. The project was built as an experiment under strict consumer hardware constraints, utilizing 24GB of unified memory and Apple’s Metal Performance Shaders. The process involved streaming data rather than downloading it, which highlighted inefficiencies quickly. The final results included training on 204.8 million tokens, with about 13 hours of pretraining and 4 minutes of supervised fine-tuning, all while consuming roughly 1 kilowatt-hour of electricity at a cost under $0.50. The choice to build a small language model stemmed from the need for precision in CLI command generation rather than creative language use. The model had to produce exact, structured commands, making it a strong test for small models under tight constraints. The training pipeline was designed with each component tailored to maximize efficiency and clarity, from the tokenizer to the evaluation process. The evaluation loop was strict and repeatable, focusing solely on exact-match accuracy, which eliminated subjective judgments about quality. While the performance of the local setup can't compete with enterprise hardware in terms of speed, it offers advantages in cost and flexibility. Training on local hardware allows for zero marginal costs on experiments and no waiting time for cloud resources. It enables developers to iterate quickly and learn from failures without incurring significant expenses or delays. The M4 Mac Mini provides a practical solution for those looking to experiment with model training without the usual financial burdens associated with cloud services.