M5 Ultra vs DGX Spark: Which Desktop Runs a 125B Model?

Apple's M5 Ultra Mac Studio tops out at 512GB of unified memory and 1.2 TB/s of bandwidth — four times the capacity of NVIDIA's DGX Spark. A Dev.to breakdown published this week argues that spec, paired with a new 125-billion-parameter Qwen model, finally makes a desktop a credible inference box.

Bandwidth is the spec that decides this

Per that breakdown, the M5 Ultra ships with a 36-core CPU, an 80-core GPU, up to 512GB of unified memory and 1.2 TB/s of memory bandwidth. The author puts that at 2.25x the DGX Spark's 273 GB/s and nearly 5x AMD's Strix Halo at 256 GB/s.

That single number carries most of the story, because token generation is memory-bandwidth-bound: every token means reading model weights out of RAM. GPU core counts are the headline; bandwidth is what you feel.

The model that makes the hardware worth owning

Alibaba's Qwen 3.8-Flash-Next landed in the same window. It is a Mixture-of-Experts model with 125 billion total parameters but only 6 billion activated per token — a 20:1 ratio between capacity and per-token compute, per the write-up. At 4-bit quantization those 125B parameters need roughly 62GB of RAM, which fits on a 128GB machine with room left over.

The practical read: the 512GB configuration is not what makes this work. A 128GB box already clears the bar for the model people actually want to run.

How the three desktops compare

Machine	Memory	Bandwidth	Base price
M5 Ultra Mac Studio	Up to 512GB unified	1.2 TB/s	$5,499
NVIDIA DGX Spark	128GB unified	273 GB/s	$4,699
AMD Strix Halo	128GB unified	256 GB/s	$3,999

Availability splits the Apple line: the write-up says the September 22 launch caps out at 256GB, and the 512GB configuration does not ship until late October. Buyers shopping for inference today are really evaluating the 192GB build at $5,499.

Where the DGX Spark still wins

Two advantages survive the bandwidth gap. On long-prompt processing, the DGX Spark is reported to run 3-4x faster prefill than comparable Mac hardware — decisive if your workload is an agent pushing 50-100K token prompts hundreds of times a session. And Blackwell has a native FP4 path that, per the same piece, Apple does not.

AMD's Strix Halo, built on the Ryzen AI MAX+ 395, undercuts both at $3,999 with 128GB and native Windows 11. Slower, but it runs the same 62GB model.

4-bit stopped being a compromise

A separate result out this week strengthens the local case. Multiverse Computing's Quantization-Aware Healing, published on Hugging Face, took a GPT-OSS 120B model compressed to 60B and quantized to MXFP4, and the 4-bit version beat its own bfloat16 checkpoint on 7 of 9 benchmarks — up 7.4 points on long-context reasoning (AA-LCR) and 5.6 on AIME 2025 math. It even edged the full-size teacher on LiveCodeBench, 66.5 to 66.0. The two losses, MMLU-Pro and SciCode, were under 1.5 points.

That matters here because every local-inference plan on this hardware runs through 4-bit weights.

What to buy

If your prompts are short and your bill is large, the bandwidth advantage is real and the Mac wins on token generation. If you run agentic coding with huge contexts, prefill speed points at the DGX Spark. If price is the binding constraint, the Strix Halo runs the same models for $1,500 less.

The open question is what arrives before late October. By the write-up's own reckoning nothing shipping today needs 512GB, and that configuration only becomes interesting if the next Qwen or DeepSeek release puts a 400B-class dense model within reach.