Hardware testing published by Tom’s Hardware pitted the new Apple Mac Studio against Nvidia DGX Spark systems to see how desktop silicon handles local artificial intelligence workloads. Reviewer Jeffrey Kampman evaluated Apple’s updated desktop with an M5 Ultra chip against Nvidia’s GB10 hardware in agentic tests and multi-billion-parameter open models [2]. The findings arrive as leaker Aaron Perris revealed details showing Apple is developing a native application named Harbor to turn desktop Macs into private inference servers [1].
Apple Mac Studio Agentic Performance
The tested Apple Mac Studio arrived with an M5 Ultra chip featuring a 36-core CPU and an 80-core GPU equipped with Neural Accelerators for matrix math operations that power local artificial intelligence jobs. The tested 256GB configuration costs $12,300. Buyers choosing maxed-out setups face shipping delays of 13 to 15 weeks from Apple, while single Dell Pro Precision GB10 systems with 48 Blackwell SMs have jumped to $10,000 during memory supply crunches. That matchup created a wide gap in memory bus speed, pitting Apple’s 1.2 TB/s against Nvidia’s 273 GB/s decode bandwidth. It gave testers an immediate look at how Apple’s unified memory bus compares with discrete accelerator architectures in day-to-day use [2].
To evaluate speed, the test used the AgentPerf-Local benchmark from Artificial Analysis to play back recorded agent task traces. Because prompt processing takes up most of this run, the test favors machines with wide memory buses and heavy compute cores over setups built primarily around rapid single-token generation. Both platforms ran dense Qwen 3.8 27B. Testers used the llama.cpp engine on both boxes with native multi-token prediction (MTP) turned on to help speed up token decoding [2].
Apple’s M5 Ultra finished in half the time. Kampman said that while paper specs do not always match real speed, the combination of compute cores and memory bus width gave Apple an unmistakable edge during these dense model runs. The $12,300 desktop easily outpaced the single $10,000 Dell GB10 box running the same inference backend [2].
Power Draw and Task Energy
Because AgentPerf-Local runs a fixed amount of work on every machine, testers were able to measure total electricity drawn straight from the wall socket to evaluate true task efficiency. Kampman converted wall draw into kilojoules of energy. The measurements showed that Apple’s M5 Ultra desktop finished faster and used roughly 16% less task energy than the DGX Spark [2].
Testers combined task time and energy use to find the energy-delay product (EDP) score for each test system. A lower EDP score indicates better overall efficiency. The Apple Mac Studio achieved an EDP score under half that of the DGX Spark, proving that Apple’s unified memory design kept efficiency high during heavy prompt processing runs [2].
While a different engine could produce better energy numbers for Nvidia chips, this head-to-head test under llama.cpp gave Apple silicon an unmistakable advantage in power draw per job. Lower power draw also kept heat down. Local developers get shorter wait times and smaller power bills from the M5 Ultra desktop build [2].
Apple Mac Studio on Heavy Models
Dense models like Qwen 3.8 27B do not stress large RAM pools, prompting testers to load bigger open-weight models including GLM-5.3 Flash and DeepSeek V4.1 Flash 1017 onto the test machines. Unsloth built the two-bit DeepSeek quant. The 256GB Mac Studio held that two-bit quantization and a four-bit build of GLM 5.3 in RAM, but a single 128GB DGX Spark could not load either model [2].
Fitting those huge models on Nvidia hardware required clustering two Spark boxes together to pool their memory into a single 256GB pool capable of holding the required weights. Dual Dell GB10 boxes cost roughly $20,000. Cheaper paired setups still run near $15,000, while Lenovo listed its rival ThinkStation PGX systems as coming soon during testing [2].
Running massive model weights on one desktop appeals to creators who followed the Apple Creator Studio AI tools update and want simple setups without clustering cables or extra power cords. A single box saves space and cash. Can raw memory capacity alone compensate when inference software lacks specific optimizations for Apple silicon [2]?
Inference Engine Differences
Software optimization proved decisive when evaluating DeepSeek V4.1 Flash 1017 between the competing setups. Due to limited software options on macOS, the Mac Studio relied on llama.cpp and could not fit both model weights and the needed DSpark speculative decoding model in RAM. Meanwhile, the Spark cluster ran the TensorFold inference engine with a custom two-Spark serving recipe created by Mia’s AI Lab that supported full speculative decoding [2].
Large context windows exposed significant throughput gaps between the two test platforms. The dual DGX Spark cluster processed the longest context lengths in under five minutes, whereas the Apple Mac Studio running unaugmented llama.cpp required between 10 and 25 minutes. Mia’s recipe gave Nvidia an easy win [2].
Results changed when testers switched to GLM 5.3 Flash, because its multi-token prediction works without needing a separate drafting model to boost token generation speeds. A TensorFold recipe built for Apple’s MLX framework helped close the prompt processing gap with the Nvidia cluster. The Mac then dominated on token decode. Memory bandwidth pushed decode speeds to three times faster on the Mac Studio, showing that software recipes make or break Apple silicon performance on big models [2].
The Harbor Thunderbolt Cluster Leak
While testers examined standalone limits, details emerged about Apple’s internal plans for multi-Mac scaling. Reporter Ryan Christoffel at 9to5Mac shared findings from leaker Aaron Perris on X detailing an unreleased Apple app named Harbor that turns Macs into private inference servers. Perris found Harbor under Claris branding. The software runs open-weight models locally on Apple silicon so that private user data never leaves local hardware or touches third-party cloud infrastructure [1].
Leaked details show Harbor automatically detects connected devices over Thunderbolt 5 cables to build an instant computer cluster without requiring manual network configuration or command line skills. Harbor pools memory across connected Macs. By sharing unified memory from two or more Macs into one pool, the app lets users run models that exceed single-machine limits and serve higher numbers of concurrent user requests. The app includes internal tools to browse, download, and load open models directly on desktop hardware. Perris reported that Harbor is designed to let users plug in and start immediately, removing the setup barriers that often complicate local distributed inference across multiple workstations [1].
Harbor is currently in private preview testing, with a public rollout expected in the coming months. Christoffel said Mac popularity for local machine learning has grown this year, aided by budget machines like the MacBook Neo and high-end hardware like the Mac mini and Mac Studio. Both models sit on Best Mac lists. If Harbor makes multi-device clustering effortless, linking Macs over Thunderbolt could soon give independent developers an accessible path to running frontier-class models at home [1].
- ONLINE NEWS Christoffel, R. (2026, November 10). Apple has new Mac app coming to make running local AI models easy. 9to5Mac. [Article Link]
- ONLINE NEWS Kampman, J. (2026, November 10). Going deep on local AI performance benchmarking with Apple’s Mac Studio with M5 Ultra taking on Nvidia DGX Spark. Tom’s Hardware. [Article Link]
