Key fact: DGX Spark: 64 or 128 GB LPDDR5x, 273 GB/s.
How we got it: NVIDIA DGX Spark spec page, checked 14 October 2026.
Checked:
DGX Spark vs Mac Studio vs Strix Halo for Local LLMs
By LLM Hardware team · Methodology ·
Key facts
- Bandwidth
- DGX Spark 273, Mac Studio 460 to 1,200, Strix Halo 256 GB/s
- Max memory
- 128 GB, 512 GB, 128 GB
- Checked
- 2026-10-13
Checked 2026-10-13. Author: LLM Hardware team. This site does not store or show Amazon prices. Check current price on Amazon.
Citeable facts
Key fact: Mac Studio: M5 Max 460 or 614 GB/s up to 128 GB; M5 Ultra up to 512 GB, 1.2 TB/s on the 80-core GPU chip.
How we got it: Apple Mac Studio specs, checked 14 October 2026.
Checked:
Key fact: Ryzen AI Max+ 395: up to 128 GB LPDDR5x-8000 on a 256-bit bus (256 GB/s).
How we got it: AMD product page, checked 14 October 2026; bandwidth computed from bus width and speed.
Checked:
Methodology — parameter math, quantization bytes, and the source list.
NVIDIA DGX Spark, Apple Mac Studio and AMD Strix Halo mini PCs all trade a discrete GPU for a large pool of unified memory. They differ most in bandwidth, maximum memory and software stack. Specs below are from the vendor pages, checked 14 October 2026. Speed figures are not published here.
Spec comparison
| DGX Spark | Mac Studio (M5 Max / M5 Ultra) | Strix Halo (Ryzen AI Max+ 395) | |
|---|---|---|---|
| Memory | 64 GB or 128 GB LPDDR5x | 36-128 GB / 96-512 GB | up to 128 GB LPDDR5x-8000 |
| Bandwidth | 273 GB/s | 460 or 614 GB/s / 1.2 TB/s (80-core GPU) | 256 GB/s |
| Software | CUDA, NVIDIA AI stack | Metal, MLX, llama.cpp, Ollama | ROCm or Vulkan, llama.cpp, Ollama |
| Vendor claim | Inference up to 200B parameters (128 GB) | n/a | Radeon 8060S, 40 GPU cores |
Model size by memory
| Memory | Largest dense at Q4 (formula) | Example |
|---|---|---|
| 64 GB | about 100B | Llama 3.3 70B Q4 (44 GB) |
| 128 GB | about 210B | Llama 3.3 70B Q8 (86 GB), gpt-oss-120b |
| 256 GB (Mac only) | about 420B | Qwen3-235B-A22B Q4 (143 GB) |
Those maxima ignore the OS share and context, so plan a few GB below them. Fit is from our VRAM formula: (memory - 2) / 0.6.
Which one to pick
- DGX Spark if your work is CUDA: fine-tuning, TensorRT-LLM, or code that has to run on NVIDIA servers later.
- Mac Studio if you want the most bandwidth or more than 128 GB. It is the only one of the three that goes to 512 GB.
- Strix Halo if you want 128 GB in an x86 Windows or Linux box and can live with 256 GB/s.
When another option is better
For models up to 32B, a 24 GB or 32 GB discrete GPU is much faster than any of these: the RTX 4090 lists 1,008 GB/s and the RTX 5090 1,792 GB/s. See best LLMs for 32 GB VRAM. For one large MoE on a GPU you already own, try MoE hybrid offloading first.
Hardware links
Check current price on Amazon. DGX Spark and larger Mac Studio configurations are sold through NVIDIA and Apple.
GMKtec EVO-X2 128 GB on Amazon
ASIN B0F53MLYQ6. Amazon Associates link. The live price is on that page, not locked in here.
Mac Studio M5 Max 36 GB on Amazon
ASIN B0HGKSQMX6. Amazon Associates link. The live price is on that page, not locked in here.
Sources
- NVIDIA, DGX Spark specs (checked 14 Oct 2026)
- Apple, Mac Studio technical specifications (checked 14 Oct 2026)
- AMD, Ryzen AI Max+ 395 specs (checked 14 Oct 2026)
- gpt-oss-120b model card (117B total, 5.1B active)
- Qwen3-235B-A22B model card (235B total, 22B active)
- LLMHardware methodology: the VRAM formula
Questions
Which is fastest for local LLMs?
By vendor bandwidth the Mac Studio leads (460 GB/s to 1.2 TB/s), then DGX Spark (273 GB/s) and Strix Halo (256 GB/s). No in-house speed tests are published here.
Which holds the biggest model?
The Mac Studio M5 Ultra, up to 512 GB. DGX Spark and Strix Halo top out at 128 GB.
A correction or a hardware question goes to the contact page.