The Apple Mac Studio with the M5 Ultra chip, featuring 256GB of memory and an 80-core GPU, was tested against Nvidia DGX Spark-based systems to evaluate local AI inference performance. The Mac Studio configuration tested is priced at approximately $12 300.
In tests using the Qwen 3.8 27B model, the Mac Studio completed tasks in under half the time of a single Dell Pro Precision GB10 system, which costs $10 000. The Mac Studio also demonstrated 16% better power efficiency and a lower energy-delay product.
For larger models like DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash, the Mac Studio utilizes its 256GB of RAM. However, the Nvidia Spark cluster, which can be scaled by pairing units, currently offers superior software support through the TensorFold engine and specialized serving recipes.
While the Mac Studio shows high performance potential, it currently faces limitations due to the maturity of the MLX stack and inference engine support compared to the Nvidia ecosystem. As of the report, lead times for the Mac Studio extend into February 2027.
