Apple's M5 Ultra Mac Studio beats a single Nvidia DGX Spark-class box by a wide margin on a small model, but a pair of Sparks wins on the large open models that justify buying 256GB of memory in the first place.
The reviewers ran Artificial Analysis' AgentPerf-Local benchmark, which replays recorded agentic AI traces and leans mostly on prompt processing. Using the dense Qwen 3.8 27B model and llama.cpp (an open-source inference engine) on both machines, the Mac finished in just under half the time. It also used about 16% less energy, and its energy-delay product, which folds speed and energy into one number, came in under half the Spark's. The Spark stand-in was a Dell Pro Precision GB10 with 4TB of storage, which the reviewers priced at $10K. The Mac has an 80-core GPU against 48 Blackwell SMs, and 1.2 TB/s of memory bandwidth against 273 GB/s.
Larger models changed the picture. Quantized versions of DeepSeek V4.1 Flash 1017 (two-bit) and GLM 5.3 Flash (four-bit) fit in the Mac's 256GB but not on a single 128GB Spark, so the reviewers clustered two Sparks. On DeepSeek, the Mac needed 10 to 25 minutes to digest the longest contexts. The Spark pair, running a community recipe from Mia's AI Lab with the TensorFold engine and DSpark speculative decoding, needed under five minutes. The Mac could not hold the DSpark drafting model alongside the weights. On GLM 5.3 Flash, the Mac fared better, but the Spark pair still processed prompts faster across the board.
Price complicates the comparison. A two-Spark setup runs $15K or more, with a dual-Dell cluster near $20K. The reviewers called that 25% above the Mac, but they used a rounded $12K; against $12,300 it is about 22%. The Mac also ships in 13 to 15 weeks, and Lenovo's ThinkStation PGX showed as "coming soon" during testing.
The pattern is that the hardware gap matters less than the software around it. The reviewers noted that Apple Silicon has few compatible inference engines, which is why llama.cpp was their only option for DeepSeek on the Mac. Whether that list grows will shape how the Mac looks on the next round of big models.