The Mac writes 2.2x faster
DeepSeek V4 Flash at 39.35 tok/s on an M5 Max against 18.05 on the Spark, on a 2K prompt with the same engine and file and no drafting. The Mac’s memory runs at 614 GB/s, the Spark’s at 273.
DGX Spark vs Mac Studio
Mac Studio vs DGX Spark for local AI models. The DGX Spark is NVIDIA’s desktop on the GB10 chip, with 128 GB of memory and CUDA; the Mac Studio is Apple’s desktop with up to 512 GB of unified memory and no CUDA. We put the M5 Max and the Spark on one matched benchmark, then set the M3 Ultra, the M4 Max and the Mac mini against the Spark on memory and bandwidth.
The Spark reads 2.1x faster
At a 64K-token prompt the Spark reads 822.98 tok/s against 398.5 on the M5 Max, which matters most when an agent or a long document fills the prompt.
Mac from $5,399
A 128 GB Mac Studio M5 Max is $5,399 on Apple’s store, a 128 GB Spark $6,950. The M5 Ultra goes to 512 GB at about 1.2 TB/s.
DGX Spark vs Mac Studio for LLMs: which is faster?
The Mac Studio M5 Max writes DeepSeek V4 Flash 2.2x faster than the DGX Spark, and the Spark reads a 64K-token prompt 2.1x faster, in antirez’s ds4 benchmark with the same engine, file and prompts on both at four prompt lengths.
Model
Measure
M5 Max writes 2.2x as fast as DGX Spark with a 2,048-token prompt and 2.4x with a 16,384-token prompt and 2.4x with a 32,768-token prompt and 2.0x with a 65,536-token prompt, same engine and file.
DGX Spark reads 1.05x as fast as M5 Max with a 2,048-token prompt and 1.5x with a 16,384-token prompt and 1.5x with a 32,768-token prompt and 2.1x with a 65,536-token prompt, same engine and file.
DGX Spark vs M3 Ultra
No matched run of the DGX Spark and the M3 Ultra exists. On paper the M3 Ultra reads memory at 819 GB/s, 3.0x the Spark’s 273 GB/s, and it launched with up to 512 GB. One LocalMaxxing user reports about 33 tok/s for DeepSeek V4 Flash on a 512 GB M3 Ultra with ds4 on Metal, against 18.05 for the Spark in the ds4 sweep, but the files and prompts differ, so read it as a rough sign.
DGX Spark vs M5 Max and M5 Ultra
The M5 Max is the one Mac with a matched run: 39.35 against 18.05 tok/s on a 2K prompt and 27.64 against 13.84 on a 64K prompt, with both machines at 128 GB. The M5 Ultra has no matched run yet; Apple lists about 1.2 TB/s, more than four times the Spark’s bandwidth, and up to 512 GB.
DGX Spark vs M4 Max
No run on the M4 Max is in our data, so no matched comparison with the DGX Spark exists. The 40-core M4 Max reads memory at 546 GB/s, 2.0x the Spark, and stops at 128 GB like the Spark, so the two hold the same models. Writing speed on these models follows bandwidth, so expect the M4 Max between the Spark and the M5 Max; we have not measured it.
DGX Spark vs Mac Studio speed test: watch them write
The Spark and the M5 Max write the same answer here at their measured ds4 speeds. If you are weighing the two after a Reddit thread, this is a matched run to quote: one engine, one file and one prompt on both.
Race
Task
Same engine, file and 2K prompt on both, no drafting (antirez ds4).
M5 Max0.0 s
writing at 39.35 tok/s
DGX Spark0.0 s
writing at 18.05 tok/s
DGX Spark and Mac Studio specs side by side
The DGX Spark’s GB10 reads memory at 273 GB/s, under half the 614 GB/s of a Mac Studio M5 Max.
| Spec | DGX Spark (GB10) | Mac Studio (M5 Max) | Mac Studio (M5 Ultra) |
|---|---|---|---|
| Memory | 128 GB LPDDR5X, or 64 GB from other brands | 36 GB on the 32-core GPU; 48, 64 or 128 GB on the 40-core | 96, 256 or 512 GB unified |
| Memory speed | 273 GB/s, 256-bit bus | 614 GB/s (460 on the 32-core GPU) | 1,228 GB/s |
| Processor and GPU | 20 Arm cores and a Blackwell GPU, 1 PFLOP at FP4 | 18 CPU cores, 32 or 40 GPU cores | 30 or 36 CPU cores, 64 or 80 GPU cores |
| Software | CUDA, DGX OS | MLX, Metal, llama.cpp | MLX, Metal, llama.cpp |
| Power | 140 W chip, 240 W adapter; 104 to 160 W at the wall under load | Up to 480 W | Up to 480 W |
| Networking | ConnectX-7, 200 Gbps, links two Sparks | 10 GbE, Wi-Fi | 10 GbE, Wi-Fi |
| Size | 150 x 150 x 50.5 mm | 95 x 197 x 197 mm, 2.7 kg | 95 x 197 x 197 mm, 3.6 kg |
| Sources | NVIDIA, ServeTheHome | Daring Fireball, MacRumors | Daring Fireball, MacRumors |
Which models fit on the DGX Spark and the Mac Studio?
The DGX Spark and a 128 GB Mac Studio M5 Max hold the same models, up to DeepSeek V4 Flash (101.5 GB). DeepSeek V4.1 Flash (about 191 GB) needs a 256 or 512 GB M5 Ultra.
| Model and size | NVIDIA DGX Spark128 GB unified | Mac Studio (M5 Max), 48 GB48 GB unified | Mac Studio (M5 Max), 128 GB128 GB unified | Mac Studio (M5 Ultra), 256 GB256 GB unified | Mac Studio (M5 Ultra), 512 GB512 GB unified |
|---|---|---|---|---|---|
| Qwen3.8-27B15.2 GB | Yes | Yes | Yes | Yes | Yes |
| DeepSeek V4 Flash101.5 GB | Yes | No | Yes | Yes | Yes |
| Qwen3.8-Flash-Next106.3 GB | Yes | No | Yes | Yes | Yes |
| Qwen3.8-27B and DeepSeek V4 Flash at once116.7 GB | Just, little room left | No | Just, little room left | Yes | Yes |
| DeepSeek V4.1 Flash190.8 GB | No | No | No | Yes | Yes |
DGX Spark vs Mac mini
The Mac mini holds at most 64 GB, on the M5 Pro chip at 307 GB/s, half the DGX Spark’s 128 GB; that 64 GB model cost $2,699 on Apple’s US store on 3 November 2026.
A 64 GB Mac mini runs Qwen3.8-27B (15.2 GB) with room to spare, and its 307 GB/s is a little above the Spark’s 273 GB/s. It cannot hold the 100B-class models: DeepSeek V4 Flash needs 101.5 GB and Qwen3.8-Flash-Next 106.3 GB, and the Spark takes both in 128 GB.
The Mac mini with the M6 chip stops at 32 GB. We know of no matched benchmark between a Mac mini and a DGX Spark.
DGX Spark and Mac Studio prices
The DGX Spark Founders Edition costs $6,950 with 128 GB (press-reported, 30 October 2026) and a 128 GB Mac Studio M5 Max $5,399 on Apple’s US store (31 October 2026).
| Machine | Memory | Bandwidth, GB/s | Software | US price |
|---|---|---|---|---|
| NVIDIA GB10 | ||||
| NVIDIA DGX Spark | 128 GB | 273 | CUDA | $6,950 to $6,995 |
| GB10 box with 64 GB (Dell, ASUS and others) | 64 GB | 273 | CUDA | $4,999 |
| Dell Pro Max with GB10 | 128 GB | 273 | CUDA | $6,593 |
| Gigabyte AI TOP ATOM | 128 GB | 273 | CUDA | $8,000 |
| ASUS Ascent GX10 | 128 GB | 273 | CUDA | $6,060 backordered |
| Lenovo ThinkStation PGX | 128 GB | 273 | CUDA | $5,859 backordered |
| Apple Mac Studio | ||||
| Mac Studio (M5 Max) | 48 or 128 GB | 614 | MLX, Metal | 48 GB: $3,099 128 GB: $5,399 |
| Mac Studio (M5 Ultra) | 96 or 256 or 512 GB | 1,228 | MLX, Metal | 96 GB: $5,499 256 GB: $10,799 512 GB: $22,299 |
| Mac Studio (M4 Max, 2025) | 128 GB | 546 | MLX, Metal | Apple no longer sells it new |
| Mac Studio (M3 Ultra, 2025) | 256 or 512 GB | 819 | MLX, Metal | Apple no longer sells it new |
US prices read on 30 and 31 October 2026.
Where the Lucebox fits beside a DGX Spark or Mac Studio
We make the Lucebox Zero 495, so read this box as ours.
Where it fits
- DeepSeek V4.1 Flash (about 191 GB) in one workstation: 33 tok/s at a 32K prompt with 192 GB in our lab. The Spark’s 128 GB cannot hold it, and on a Mac it takes a $10,799 M5 Ultra.
- Qwen3.8-27B on its own 32 GB R9700 card at 640 GB/s, faster memory than the Spark or the M5 Max, with the unified memory left for a second model.
- Eight image questions on Qwen3.8-27B in 10.5 s against 25.4 s on a DGX Spark, measured on the R9700 in our earlier 395 build.
Where it doesn't
- You want the smallest, quietest desk machine: the Mac Studio wins there.
- Your software needs CUDA: of the three, only the DGX Spark runs it.
- You need it on your desk this month.
From $5,999, delivered from January 2027. See the hardware Configure and order
More local AI hardware comparisons
NVIDIA DGX Spark
AMD Ryzen AI MAX+ boxes
- DGX Spark vs Strix Halo Close on memory speed, about half the price, CUDA only on the Spark.
- Framework Desktop The same chip as the cheaper boxes, in a small case you can build.
- GMKtec EVO-X2 AI mini PC One of the cheapest 128 GB Strix Halo boxes, with 64 GB as an option.
- Minisforum MS-S1 MAX Strix Halo with a PCIe slot, and a 192 GB Gorgon Halo version.
- Beelink GTR9 Pro A 128 GB Strix Halo box with two 10 GbE ports.
- Corsair AI Workstation 300 Strix Halo from a big US brand, in stock.
DGX Spark vs Mac Studio FAQ
DGX Spark or Mac Studio: which should I buy?
Buy the DGX Spark if your work needs CUDA or long prompts: it reads a 64K-token prompt 2.1x faster than a Mac Studio M5 Max in antirez’s ds4 benchmark. Buy the Mac Studio if you mostly chat and write: the M5 Max writes DeepSeek V4 Flash 2.2x faster and costs $5,399 with 128 GB, against $6,950 for the Spark.
Is the Mac Studio faster than the DGX Spark for coding agents?
Usually not: coding agents resend long prompts, and the DGX Spark read a 64K-token prompt at 822.98 tok/s against 398.5 on a Mac Studio M5 Max in the ds4 sweep. The Mac still writes each reply 2.2x faster, so short back-and-forth feels quicker on it.
Can a Mac Studio run DeepSeek V4.1 Flash?
A Mac Studio M5 Ultra with 256 or 512 GB holds DeepSeek V4.1 Flash at 2 bits (about 191 GB), which a 128 GB DGX Spark cannot. How it runs on 128 and 192 GB: DeepSeek V4.1 Flash on the Lucebox.
Does the Mac Studio run CUDA?
No. The Mac Studio runs MLX, Metal and llama.cpp, and macOS has no CUDA, so CUDA-only tools need the DGX Spark or an NVIDIA card in a PC. Since April 2026 tiny corp’s TinyGPU driver lets tinygrad use an NVIDIA card over Thunderbolt on a Mac for AI compute, still without CUDA: NVIDIA eGPU on a Mac, benchmarked.