Mac Studio, DGX Spark or Strix Halo: which local-AI box should you buy?

Short answer: buy the Mac Studio if you want the fastest replies from big local models, the DGX Spark if you need NVIDIA's CUDA stack, fast prompt processing or image and video generation, and a Strix Halo mini PC if you want 128GB of memory for the least money. The deciding number for reply speed is memory bandwidth: 614 GB/s on the 40-core M5 Max Mac Studio, 273 GB/s on the DGX Spark and 256 GB/s on Strix Halo, per Apple's, NVIDIA's and AMD's own spec pages read November 3, 2026.

Which one should you buy?

Pick by the job you'll give it most, not by one spec. Here's how the three split, from vendor specs and the benchmarks below.

How much does each one cost?

A 128GB Strix Halo box costs about $3,450 to $3,500, a 128GB DGX Spark about $4,700 to $7,000 depending on where you buy, and a 128GB Mac Studio $5,399. Memory prices have been moving all year, so each price below carries the date we read it.

We could not load NVIDIA's own store on November 3, so we can't confirm today's 128GB Spark price there. The two dated figures we have point in different directions: NVIDIA's developer forum post of February 25, 2026 says the Founders Edition MSRP "has been adjusted from $3,999 to $4,699 due to memory supply constraints", and HotHardware reported on October 2 that the 128GB model's price "has increased significantly to $6,995". Check the price at checkout, and look at partner GB10 boxes (Dell Pro Max, ASUS Ascent GX10, HP ZGX Nano and others): it's the same chip.

Which one writes replies fastest?

The Mac Studio, by a wide margin. Writing a reply (decode) means reading the model's active weights from memory for every token, so speed follows memory bandwidth. Tom's Hardware put it plainly in its July 30 test: "the speed of the entire decode process basically becomes dependent on how fast those model weights can be streamed in from GPU memory."

The Strix Halo figure is arithmetic: AMD lists 256-bit LPDDR5x-8000, which works out to 256 GB/s. Tom's Hardware also describes the M4 Max as having "double the memory bandwidth of GB10 and Strix Halo on paper". (An earlier version of this page said 273 GB/s for Strix Halo; that is GB10's number.)

What the three-way benchmark measured

Tom's Hardware tested a 128GB M4 Max Mac Studio (546 GB/s, the previous generation) against GB10 and Strix Halo systems with llama.cpp and four-bit Unsloth quantizations. The Mac generated more tokens per second than both at every context depth, but its lead varied a lot by model:

Against Strix Halo, the same Mac was 2.26x faster on Gemma 4 12B and 1.6x faster on gpt-oss-120b. Tom's summed it up: "you might see the M4 Max generate as few as 27% more tokens per second or as many as 82% more TPS than an Nvidia GB10 box." Twice the bandwidth did not mean twice the speed on mixture-of-experts models, which read only part of their weights per token.

The current Mac Studio is faster still. In its September 21 review of the M5 Ultra Mac Studio, Tom's Hardware wrote: "The M5 Ultra's prompt processing speeds are even faster than Nvidia's DGX Spark's, and its tokens-per-second throughput is double that of the M4 Max and almost four times higher than the DGX Spark across the board." Nobody we found has published the same three-way test on the M5 Max yet; its 614 GB/s sits about 12% above the M4 Max's 546 GB/s, so expect it to lead the Spark by at least as much as the M4 Max did.

Which one reads long prompts fastest?

The DGX Spark. Reading your prompt (prefill) is compute-bound, not bandwidth-bound, and GB10 has the most GPU compute of the three. In Tom's tests the M4 Max landed "behind GB10, especially with longer contexts" on Qwen 3.6, and on Gemma 4 12B its prompt speed was "only in line with (or even slightly worse than) Strix Halo."

That matters for agents, coding tools and retrieval, where the prompt is long and the reply short. In Tom's fixed gpt-oss-120b job (a 2,048-token prompt on a 32,768-token context, then 1,024 output tokens), the GB10 finished faster and used less total energy than the Mac Studio, "even though it's outputting fewer tokens per second in the decode stage." The M5 Ultra changes this, per the quote above, but at $5,499 and up it's a different budget. Our DGX Spark review has llama.cpp's own prefill and decode numbers for the Spark.

Which one is best for image and video generation?

The DGX Spark. Diffusion models are bound by GPU compute, and Tom's measured the M4 Max taking longer than even Strix Halo's Radeon 8060S on its Flux.2 Klein test. The Mac's first run failed outright, from "an apparent lack of GPU support for the FP8 data type that the ComfyUI workflow uses by default," and only ran after switching to the FP16 model. Apple claims the M5 Max is "up to 3.8x faster" at image generation than the M4 Max, which Tom's quotes but couldn't test. If ComfyUI is your main use, CUDA is the safer bet today.

How big a model can each one run?

All three run 120B-class mixture-of-experts models at 4-bit in 128GB. Only the Mac Studio M5 Ultra goes past 128GB in one box.

• DGX Spark: NVIDIA lists one 128GB unit for models up to 200B parameters and two linked over ConnectX-7 for up to 400B; its blog says the 64GB version "supports up to 100-billion-parameter models".
• Strix Halo: on a 128GB Ryzen AI Max+ 395, AMD's Variable Graphics Memory gives the GPU up to 96GB, and AMD says the GPU "can technically access a total graphics memory size of 112GB". Details in our Strix Halo VRAM guide.
• Mac Studio: M5 Max tops out at 128GB; M5 Ultra offers 96GB, 256GB or 512GB per Apple's spec page. macOS keeps some memory for itself, so don't plan on using every gigabyte for the model.

Check your exact model and quantization in our VRAM calculator before you buy; a 70B dense model at 4-bit fits on all three but generates slowly on the 256 to 273 GB/s machines. For gpt-oss-120b specifically, see the gpt-oss-120b hardware guide.

Should you rent one before you buy?

If you're on the fence about the Spark, yes: Vast lists GB10 machines by the hour. On November 3, 2026 at 20:20 UTC our price tracker read the cheapest single GB10 on Vast at $0.28 an hour on demand, so a few evenings of testing your own models costs a few dollars against a $4,999-plus purchase. Prices move hourly; our cloud GPU page has today's.

Apple Silicon and Strix Halo aren't rentable in the same way, so for those the published benchmarks above are what you have. Our rent vs buy guide works through when owning pays off.

What we did not test

We have not run these machines ourselves. Speeds here are Tom's Hardware's measurements (M4 Max, GB10 and Strix Halo in July 2026; M5 Ultra in September 2026), quoted word for word from their pages, and the M5 Max comparison is our inference from Apple's bandwidth figure, labeled as such. Prices are what each vendor's page showed on November 3, 2026, except the 128GB Spark, explained above. No referral fee depends on which machine you pick; the only referral link on this page is the Vast rental card.

The simple rule: if you'll mostly chat with or code against a big model, the Mac Studio's bandwidth wins. If you'll feed it long prompts, run image models or need CUDA, the Spark is the better tool. If you want the most memory for the least money and can live with slower replies, buy Strix Halo. We'll update this page when a three-way test of the M5 Max or the new Ryzen AI Max+ PRO 495 (Framework lists a 192GB version at $6,799 on pre-order, shipping in December) is published.