Updated October 2026
DGX Spark vs Mac Studio vs Strix Halo: the three 128GB desktop boxes
Three small, quiet desktops hold a 70B model in memory. They differ by more than double in speed, by a lot in software, and by a lot in price. Which one for running models, which for building on them, and which to skip.
As an Amazon Associate I earn from qualifying purchases. Product links on this page go to Amazon; they are affiliate links, and the price you pay is the same.
The picks
GMKtec EVO-X2 (Ryzen AI Max+ 395, 128GB)
The cheapest way to hold a 70B model. Slowest of the three at generating, mature x86 software, Linux or Windows. Buy it to run big models, not to run them fast.
Up to 96GB can be assigned to the GPU on Windows, more on Linux. That loads 70B models on a machine that costs less than one 5090. 256GB/s of bandwidth means they run, not race.
Mac Studio (M5 Max, 128GB unified memory)
Twice the bandwidth of the other two, so twice the tokens per second on the same model. The one to buy if you'll use the models rather than develop on them.
The smallest quiet box that runs 70B models at a speed you'd use every day. The 128GB Strix Halo boxes fit the same models at less than half the bandwidth.
NVIDIA DGX Spark (GB10, 128GB)
The same memory and about the same speed as Strix Halo, plus CUDA and NVIDIA's full stack. Buy it to fine-tune, build and deploy; it earns its price only if you'll use that stack.
If you're fine-tuning, not just running, this is the smallest machine that does it the way the datacenter does.
What they have in common
All three put 128GB of memory where both the CPU and GPU can reach it, in a box that fits on a shelf and runs quietly. That’s enough to load a 70B model at 4-bit (about 42GB), gpt-oss-120B (about 63GB), or a 235B mixture-of-experts at Q3. No consumer graphics card does that; the biggest, the RTX 5090, has 32GB.
Past that, they’re different machines.
The numbers
| GMKtec EVO-X2 (Strix Halo) | NVIDIA DGX Spark | Mac Studio M5 Max | |
|---|---|---|---|
| Chip | AMD Ryzen AI Max+ 395 | NVIDIA GB10 (Grace + Blackwell) | Apple M5 Max, 40-core GPU |
| Memory | 128GB LPDDR5X | 128GB LPDDR5X | 128GB unified |
| Memory bandwidth | 256GB/s | 273GB/s | 614GB/s |
| Max memory for the GPU | ~96GB (Windows), more on Linux | ~120GB | ~110GB after macOS |
| gpt-oss-120B generation | ~55 t/s (community, llama.cpp) | 59 t/s (measured, llama.cpp) | ~120 t/s (estimated from bandwidth) |
| 70B dense at Q4, generation | ~4 t/s | ~4.5 t/s | ~10 t/s |
| Prompt processing | weakest | strong (Blackwell, FP4) | good |
| Software | x86 Linux or Windows; llama.cpp (Vulkan, ROCm), Ollama, LM Studio | NVIDIA’s Linux; full CUDA stack, vLLM, TensorRT, fine-tuning tools | macOS; MLX, llama.cpp Metal, Ollama, LM Studio |
| Also sold as | Beelink GTR9 Pro, Framework Desktop (direct only), HP Z2 Mini G1a | ASUS Ascent GX10, Dell, HP, Lenovo variants | Mac Studio M5 Ultra (256/512GB, 1.2TB/s) |
| Price tier | $1,000 to $3,000 | No ceiling | No ceiling (128GB is configure-to-order) |
Bandwidth decides generation speed on every one of them, which is why the Mac generates at roughly double the other two on the same model. The DGX Spark’s 59 tokens a second on gpt-oss-120B is a real llama.cpp measurement; the Strix Halo figure is from community benchmarks on the same model; the Mac figure is scaled from bandwidth and should be treated as an estimate until someone publishes the M5 Max number. A dense 70B model is slow on all three; the mixture-of-experts models are what make these boxes feel fast.
Pick by what you’ll do
You want to run big models and use them. Chat, agents, coding assistants, long documents. The Mac Studio M5 Max. Double the speed is the difference between 10 tokens a second and 5 on a 70B model, and between a tool you use and one you demo. The software is mature and nothing needs configuring. It costs the most and it’s worth it for this job.
You want to hold big models for the least money. Run gpt-oss-120B, try 70B models, serve a small team over the network, and you don’t care that it’s slower. The EVO-X2. It’s the only one of the three in the middle tier, it runs the same x86 software as any PC, and ROCm on this chip has been good. Accept that 70B dense models are a 4-tokens-a-second experience.
You want to build on models, not just run them. Fine-tuning, quantizing, vLLM, TensorRT, anything that assumes CUDA. The DGX Spark. It’s the same memory and roughly the same generation speed as Strix Halo at a much higher price, and the entire difference is NVIDIA’s software stack. If you’ll use that stack, nothing else on this page does the job. If you won’t, you’re paying double for a slower Mac.
You want more than 128GB. Only one answer: the Mac Studio M5 Ultra at 256GB or 512GB, 1.2TB/s. The 235B and 400B-class models live there and nowhere else on a desk.
What none of them replace
A PC with a 5090 generates on a 32B model at about 60 tokens a second, three times the Mac Studio and six times the other two, because 1.8TB/s of GDDR7 beats any shared-memory design. If the models you run fit in 32GB, buy the card, not the box. These three exist for the models that don’t fit.
Sources
- llamaperf: RTX Spark for local LLMs, what DGX Spark tells you (DGX Spark llama.cpp measurements)
- Apple: Mac Studio with M5 Max and M5 Ultra
- Beelink GTR9 Pro vs GMKtec EVO-X2