At 128 GB, DGX Spark costs less. The Mac Studio moves memory more than twice as fast.

Running an AI model locally means it runs on a computer you own, not on another company's servers. Your prompts and documents stay on the machine, it works without an internet connection, and nobody charges you per use. The catch is hardware: the model has to fit in memory, and data has to move through that memory quickly enough to be useful.

That is why this report looks at three small desktops built around unified memory, a single pool that the processor and the graphics chip share. On most PCs the graphics card has its own, separate memory; on these three the graphics chip draws on the shared pool, which is how a box that fits on a desk can hold a model tens of gigabytes in size.

Why these three? All are small computers that sit on a desk and connect to a screen, all share one memory pool between processor and graphics chip, and their prices overlap: DGX Spark's USD $4,699 falls between the cheapest and the most expensive Mac Studio, and the top Mac mini costs more than the cheapest Mac Studio. Laptops, all-in-one Macs and PCs with separate graphics cards are different comparisons and are left out.

Memory is the hard limit. A Mac mini tops out at 64 GB, and only that largest configuration has room for a 70B model. No Mac mini has room for gpt-oss-120b, so for that you need DGX Spark or a Mac Studio with at least 96 GB. Past 128 GB, the only single machine here is the Mac Studio with M5 Ultra.

Each dot is a memory option you can order, on a scale that doubles at every step. Lines A, B and C are model files: a machine needs a dot comfortably to the right of a line.

Every word a model writes means reading its weights out of memory again, so the speed memory moves at sets a ceiling on writing speed. DGX Spark's 273 GB/s is a little below the M5 Pro Mac mini's 307 and well below every Mac Studio, which run from 460 to 1,200 GB/s. On paper, DGX Spark holds as much as the 128 GB Mac Studio but writes more like a Mac mini. These are the vendors' own figures, not measurements.

Nobody has published benchmarks for the M5 and M6 Macs yet. These results are for the Macs they replace (M3 Ultra and M4 Pro) and for DGX Spark or a machine built on its GB10 chip. They show the size of the gap, not what the new Macs will do.

The measured results agree with the bandwidth figures on writing and complicate the rest. In all three matched tests, the previous Mac Studio wrote faster than the GB10 machine. Reading a prompt is split: the GB10 machine was faster with gpt-oss-20b and Llama 3.1 70B, the Mac Studio with gpt-oss-120b. These are the Macs the new ones replace, and software versions differ, so read the gaps, not the exact numbers.

At every memory size both sell, a Mac mini costs less than a Mac Studio. At 128 GB, DGX Spark's USD $4,699 is USD $400 less than the cheapest 128 GB Mac Studio, and it comes with 4 TB of storage against the Mac's 512 GB. Past 128 GB, only the Mac Studio remains: USD $9,499 for 256 GB.

The cheapest US price at each memory size, for each product. Storage differs: the cheapest Mac mini has 256 GB, the Mac Studio 512 GB or 1 TB, and DGX Spark comes only with 4 TB.

The common local tools run on both. DGX Spark adds NVIDIA's serving software and native 4-bit FP4 and 8-bit FP8 formats; MLX, Apple's own framework, is documented for the Mac only.

If your tools are llama.cpp, Ollama or LM Studio, all three machines run them. The differences are at the edges: NVIDIA's own serving software, TensorRT-LLM, does not run on a Mac; vLLM and SGLang are only partly supported there; and Apple's MLX has no official DGX Spark documentation.

Apple states the most power each Mac draws continuously. NVIDIA states the rating of DGX Spark's power supply and the chip's power limit. Those are different measurements, so they are listed, not ranked.

The prices above are US prices, because the US is the only one of our eight markets where DGX Spark has a published price and is in stock. Elsewhere, on 14 September 2026:

NVIDIA documents linking two DGX Sparks over their 200 Gb/s ConnectX-7 ports and says a pair runs models of up to 405 billion parameters. Its product page says up to four run models of up to 700 billion.

Apple added RDMA over Thunderbolt 5 in macOS 26.2, for running a model across several Macs with MLX. Apple gives no figure for how large a model that allows.

If you mostly run models up to about 20 billion parameters, or want to start small, the Mac mini is the cheapest way in: the Mac mini with M6 and 16 GB at USD $899 has room for gpt-oss-20b, and the Mac mini with M5 Pro and 64 GB at USD $2,699 for a 70B model.

If you want the largest models or the fastest writing, look at the Mac Studio. It is the only machine here that goes past 128 GB and it has the highest memory bandwidth, and the previous generation already wrote faster than a GB10 machine in every matched test.

If your work depends on NVIDIA's software or its 4-bit FP4 format, if prompt-reading speed matters most (the GB10 machine led in two of the three tests), or if you want 128 GB for the least money, DGX Spark fits better. It costs USD $400 less than a 128 GB Mac Studio and adds 4 TB of storage, but on paper it writes more slowly.

No benchmarks for the new Macs had been published when this report was written, and the gaps above may change once they are.