
Mac Studio M5 Ultra vs DGX Spark: Local AI Speed Tested
In Tom's Hardware tests, the M5 Ultra Mac Studio finished an agentic AI benchmark in under half a DGX Spark's time using 16% less energy. Full breakdown.
Unified memory machines have become the go-to hardware for running big language models at home, and the two names everyone compares are Apple's Mac Studio and NVIDIA's DGX Spark. On November 10, 2026, Tom's Hardware published a deep local AI benchmark pitting a 256GB Mac Studio with the M5 Ultra against GB10-based DGX Spark-class systems, including a two-box cluster. The verdict is nuanced, and genuinely useful if you are shopping for a local LLM workstation.
- Agentic test: on Artificial Analysis' AgentPerf-Local with Qwen 3.8 27B and llama.cpp, the Mac Studio finished in just under half the time of a single GB10 system and used about 16% less energy
- Hardware: M5 Ultra offers a 36-core CPU, 80-core GPU, up to 512GB of unified memory and 1.2TB/s of bandwidth, per Apple
- Big models: a two-Spark cluster running the TensorFold engine was much faster on DeepSeek V4.1 Flash, while TensorFold on the Mac tripled average decode speed for GLM 5.3 Flash versus llama.cpp
- Prices: Tom's tested a Mac Studio configuration around $12,300 and says a 4TB Dell Pro Precision GB10 now costs about $10,000
What Did Tom's Hardware Test?
The team used a maxed-out M5 Ultra Mac Studio with 256GB of memory and compared it against a Dell Pro Precision GB10, which Tom's describes as nearly identical to the DGX Spark. Single-box tests used the dense Qwen 3.8 27B model with multi-token prediction enabled on both platforms. For larger models, Tom's paired two GB10 boxes to match the Mac's memory pool, since neither DeepSeek V4.1 Flash nor GLM 5.3 Flash quantizations fit on a single 128GB Spark.
Is the Mac Studio Faster Than a DGX Spark for Local LLMs?
For the single-box agentic workload, clearly yes. The M5 Ultra's 80-core GPU with Neural Accelerators and roughly four times the memory bandwidth of GB10's 273GB/s gave it a commanding lead, and its energy-delay product came in at under half the Spark's. That matches Apple's pitch at the August launch, which framed the M5 Ultra Mac Studio around running frontier-class models entirely on device.
The picture shifts with the largest models. On DeepSeek V4.1 Flash, the Mac was limited to llama.cpp and couldn't fit the speculative decoding model alongside the weights, so it needed 10 to 25 minutes to digest the longest contexts Tom's tested, versus under five minutes for the Spark cluster on its optimized recipe. With GLM 5.3 Flash, however, an MLX-specific TensorFold recipe brought the Mac's prompt processing close to the cluster and tripled its decode throughput.
Why Software Recipes Matter More Than Specs
The biggest takeaway is that raw bandwidth is only half the story. Tom's concludes that the NVIDIA ecosystem's breadth of community serving recipes almost guarantees a fast way to run new open-weight models soon after release, while Apple's MLX stack is improving quickly but still catching up. If you already run models on a Mac, our guide to running local LLMs with MLX covers the basics, and our DGX Spark 64GB explainer covers the other side.
Which Local AI Box Should You Buy?
If you want one quiet, efficient desktop that is also a top-tier general workstation, the Mac Studio is a remarkable machine, as long as you are comfortable waiting for MLX recipes for the newest models. Tom's notes lead times stretching into February 2027. If your priority is running the hottest new open model the week it lands, a Spark cluster's software support is the safer bet, though two boxes with 4TB drives now run roughly $15,000 to $20,000. Explore more options in our mini PC and local AI hardware coverage.
Sources: Tom's Hardware — November 10, 2026; Apple Newsroom: Mac Studio with M5 Max and M5 Ultra — August 25, 2026.
More Mini Computers Stories

ESP32-P4 Desktop OS: Sylphide OS and BMDS Turn MCUs Into PCs
Two open-source desktop OS projects, Sylphide OS and BMDS, give the $25 Olimex ESP32-P4-PC windows, apps, SSH and even Doom at 1280x720 and 60 FPS.

Tektronix 1 Series Oscilloscope: $450 Scope With AI Control
Tektronix's 1 Series DPO starts at $450 with 12-bit resolution, up to 200 MHz, I2C/SPI/UART decoding and an AI connector for natural-language control.

Eight Used V100 GPUs: How a Lawyer Built a Private AI Server
A lawyer's private AI server runs DeepSeek V4.1 Flash on eight used V100 GPUs at up to 215 tokens per second, keeping case files off the cloud.
