Mac Studio challenges DGX Spark, but software determines the winner
Apple’s M5 Ultra Mac Studio outperformed a DGX Spark-like Dell system in an agentic benchmark, but larger models exposed significant differences in software support. Using llama.cpp with Qwen 3.8 27B, the Mac completed AgentPerf-Local in just under half the time and consumed about 16% less task energy. The tested Mac configuration cost $12,300, while the Dell system’s price had climbed to $10K.
The Mac’s 256GB of unified memory accommodated quantized versions of DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash, whereas Nvidia required a paired system configuration. Despite that capacity advantage, DeepSeek ran substantially slower on the Mac using llama.cpp. The Spark cluster benefited from TensorFold and DSpark speculative decoding, while the Mac lacked enough memory to hold both the main model and its speculative decoding model.
Software optimization substantially changed the results for GLM 5.3 Flash. Switching the Mac from llama.cpp to a TensorFold serving recipe brought prompt processing closer to the Spark cluster and made decode throughput three times faster, on average. Apple’s hardware offers compelling workstation versatility and inference potential, but performance depends heavily on model-specific support. Nvidia’s more mature software ecosystem and clustering capabilities make paired Sparks a stronger immediate option for dedicated local inference, despite their higher cost.