Strix Halo costs about half
A 128 GB Strix Halo box starts at $3,500, NVIDIA’s own 128 GB Spark is reported at $6,950. A 192 GB Gorgon Halo box is $6,799.
DGX Spark vs Strix Halo
NVIDIA’s GB10 against AMD’s Strix Halo and Gorgon Halo, chip to chip: specs, CUDA or ROCm, what fits, measured speed and the price of every box.
Close on memory speed
256 GB/s on Strix Halo, 273 on the Spark and Gorgon Halo. With drafting, our lab measured 41.9 tok/s on DeepSeek V4 Flash on a Strix Halo box; 35.3 is published for one Spark, each with its own setup.
192 GB only on AMD
Gorgon Halo boxes, AMD’s own Ryzen AI Halo among them, go to 192 GB. Every GB10 box stops at 128 GB.
DGX Spark vs Ryzen AI MAX+ 395: specs on paper
Strix Halo is AMD’s Ryzen AI MAX+ 395, a chip with up to 128 GB of unified memory at 256 GB/s, and the DGX Spark’s GB10 has 128 GB at 273 GB/s, so on paper the two read memory at close speeds. The GB10, the 395 and the PRO 495 all use a 256-bit memory bus; Gorgon Halo adds faster memory and room for 192 GB.
| Spec | DGX Spark (GB10) | Strix Halo (Ryzen AI MAX+ 395) | Gorgon Halo (Ryzen AI MAX+ PRO 495) | Lucebox Zero 495 |
|---|---|---|---|---|
| Memory | 128 GB LPDDR5X, or 64 GB from other brands | Up to 128 GB LPDDR5X-8000 | Up to 192 GB LPDDR5X-8533 | 32 GB on the R9700 card, plus 128 or 192 GB unified |
| Memory speed | 273 GB/s, 256-bit bus | 256 GB/s, 256-bit bus | 273 GB/s, 256-bit bus | 640 GB/s on the card, 273 GB/s unified |
| Processor and GPU | 20 Arm cores and a Blackwell GPU, 1 PFLOP at FP4 | 16 Zen 5 cores, Radeon 8060S GPU | 16 Zen 5 cores, 40-CU Radeon GPU | Ryzen AI MAX+ PRO 495 and a Radeon AI PRO R9700 |
| Software | CUDA, DGX OS | ROCm, Vulkan | ROCm, Vulkan | ROCm, Vulkan, Lucebox engine |
| Power | 140 W chip, 240 W adapter; 104 to 160 W at the wall under load | 120 W chip, 140 W peak (Beelink GTR9 Pro) | 120 W sustained, 140 W boost; 400 W power supply inside (Framework Desktop) | 1,000 W power supply, SFX form factor |
| Networking | ConnectX-7, 200 Gbps, links two Sparks | Depends on the box | Depends on the box | 10 GbE, Wi-Fi 7, optional 100 GbE cluster kit |
| Size | 150 x 150 x 50.5 mm | About 4.5 L (Framework Desktop) | About 4.5 L (Framework Desktop) | 340 × 110 × 320 mm, 11.97 L, 6 kg |
| Sources | NVIDIA, ServeTheHome | Beelink, Framework | Phoronix, Framework | Lucebox |
CUDA vs ROCm: does Strix Halo run CUDA?
No, Strix Halo does not run CUDA: AMD’s Ryzen AI MAX+ chips run ROCm and Vulkan, and CUDA stays on NVIDIA hardware such as the DGX Spark.
On the Spark, NVIDIA publishes ready playbooks for vLLM, SGLang, llama.cpp, LM Studio and NIM, and for fine-tuning with NeMo, Unsloth and LLaMA-Factory. Tools that ship only for CUDA run there as they do on any NVIDIA card.
On Strix Halo, llama.cpp runs on ROCm or Vulkan, and our own Strix Halo run on this page used the Lucebox engine on ROCm. In our lab the Lucebox engine on ROCm read prompts faster than llama.cpp on Vulkan at every length on the same box: ROCm vs Vulkan on Strix Halo has the numbers.
A fine-tuning script or serving engine with no AMD build will not run on a Strix Halo box. If your stack is CUDA, buy the Spark; for inference with llama.cpp or a ROCm engine, either chip works.
Which models fit on a DGX Spark and a Strix Halo box?
A 128 GB DGX Spark or Strix Halo box holds every model here except DeepSeek V4.1 Flash, which needs about 191 GB and so a 192 GB Gorgon Halo box. DeepSeek V4 Flash takes 101.5 GB of the 128, and DeepSeek V4 Flash on Strix Halo covers that run in full.
| Model and size | NVIDIA DGX Spark128 GB unified | Minisforum MS-S1 MAX, 128 GB128 GB unified | Framework Desktop (Ryzen AI MAX+ PRO 495)192 GB unified | Lucebox Zero 495, 128 GB32 GB card + 128 GB |
|---|---|---|---|---|
| Qwen3.8-27B15.2 GB | Yes | Yes | Yes | Yes, on the card |
| DeepSeek V4 Flash101.5 GB | Yes | Yes | Yes | Yes, card and unified |
| Qwen3.8-Flash-Next106.3 GB | Yes | Yes | Yes | Yes, card and unified |
| Qwen3.8-27B and DeepSeek V4 Flash at once116.7 GB | Just, little room left | Just, little room left | Yes | Yes, card and unified |
| DeepSeek V4.1 Flash190.8 GB | No | No | Just, little room left | Yes, part on the SSD |
Strix Halo vs DGX Spark for LLMs: how fast does each one run?
Strix Halo writes the big MoE models about as fast as a DGX Spark: 41.9 tok/s on DeepSeek V4 Flash with drafting in our lab, against 35.3 in a published one-Spark benchmark. No one has run both chips on the same engine and file yet, so each run below names its own setup.
Model
Measure
These runs differ in engine, file or prompt.
These runs differ in engine, file or prompt. Drafting speeds some prompts far more than others.
These runs differ in engine, file or prompt.
These runs differ in engine, file or prompt. Drafting speeds some prompts far more than others.
R9700 (Lucebox card) writes 2.9x as fast as Strix Halo with a 27-token prompt, same engine and file. Drafting speeds some prompts far more than others.
These runs differ in engine, file or prompt. Drafting speeds some prompts far more than others.
These runs differ in engine, file or prompt.
Writing speed with drafting on
With a draft model, a Strix Halo box wrote DeepSeek V4 Flash at 41.9 tok/s in our lab (Lucebox engine on ROCm, 8K prompt) and Qwen3.8-Flash-Next at 51.39 tok/s in a LocalMaxxing run by soulrider4ever. On the Spark side the drafted run is an independent one-Spark reproduction with a DSpark draft, and we found no drafted DGX Spark run on Qwen3.8-Flash-Next.
Writing speed without drafting
Without drafting, schuttdev’s LocalMaxxing run had a Strix Halo box writing DeepSeek V4 Flash at 30.75 tok/s with hipfire and a 128-token context, close to a best case. antirez’s ds4 baseline puts one DGX Spark at 18.05 tok/s with a 2K prompt and 13.84 at 64K. Engine, file and prompt all differ, so the two do not give a fair ratio.
Reading long prompts
Prompt reading leans on compute, and the numbers we have favour the Spark’s Blackwell GPU: ds4’s baseline has it reading DeepSeek V4 Flash at 825.76 tok/s on a 2K prompt and 822.98 at 64K. We have no matched Strix Halo run for that model; on the larger DeepSeek V4.1 Flash, Framework measured its 192 GB Gorgon Halo box at 340 tok/s on a 32K prompt.
Qwen3.8-27B: where a graphics card wins
For Qwen3.8-27B, a Strix Halo box wrote 104.32 tok/s with drafting in schuttdev’s LocalMaxxing run, and a Radeon AI PRO R9700 card gave 303.38 on the same hipfire build, file and prompt, 2.9x faster (2.6x without drafting). We have no DGX Spark run for this model that we can show yet.
DGX Spark vs Strix Halo speed test: watch them write
Each machine’s best published setup with drafting, then the same chip with and without a card.
Race
Task
Each machine's own best published setup, all with drafting on. Engines and files differ.
Lucebox (395 build)0.0 s
writing at 51.1 tok/s
Strix Halo box0.0 s
writing at 41.9 tok/s
DGX Spark0.0 s
writing at 35.3 tok/s
Strix Halo and DGX Spark prices, box by box
A 128 GB Strix Halo box starts at $3,500 (GMKtec EVO-X2), against $5,859 for the cheapest 128 GB GB10 box we read and $6,950 reported for NVIDIA’s own Spark.
| 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 |
| AMD Ryzen AI MAX+ | ||||
| Framework Desktop (Ryzen AI MAX+ 395) DIY edition, no drive in the price. | 64 or 128 GB | 256 | ROCm, Vulkan | 64 GB: $1,959 out of stock 128 GB: $3,449 out of stock |
| Minisforum MS-S1 MAX | 64 or 128 GB | 256 | ROCm, Vulkan | 64 GB: $2,599 in stock 128 GB: $3,919 backordered |
| GMKtec EVO-X2 | 64 or 128 GB | 256 | ROCm, Vulkan | 64 GB: $2,200 in stock 128 GB: $3,500 in stock |
| Corsair AI Workstation 300 | 128 GB | 256 | ROCm, Vulkan | $4,600 in stock |
| Beelink GTR9 Pro | 128 GB | 256 | ROCm, Vulkan | $4,349 backordered |
| Framework Desktop (Ryzen AI MAX+ PRO 495) DIY edition, no drive in the price. | 192 GB | 273 | ROCm, Vulkan | $6,799 out of stock |
| Minisforum MS-S1 MAX-P495 A sale price; $9,249 regular. | 192 GB | 273 | ROCm, Vulkan | $7,399 out of stock |
| iBUYPOWER mini PC (Ryzen AI MAX+ PRO 495) | 128 GB | 273 | ROCm, Vulkan | $3,999 |
US prices read on 30 and 31 October 2026.
Ryzen AI Halo vs DGX Spark
Ryzen AI Halo is AMD’s own developer mini PC, and its Ryzen AI MAX+ PRO 495 version has 192 GB of memory at 273 GB/s, against 128 GB at 273 GB/s in NVIDIA’s DGX Spark.
Press reports put the 192 GB Ryzen AI Halo at $7,299.99 for the Linux model when it launched on 12 and 13 October 2026 (guru3d, 13 October 2026), against $6,950 reported for the 128 GB Spark. Both are mini PCs about 150 mm square, each sold by the company that makes its chip.
It ships with ROCm, and AMD has promised a Windows version later. The first Ryzen AI Halo, with the Strix Halo chip and 128 GB, was announced in May 2026 at $3,999, per Liliputing.
We have not measured a Ryzen AI Halo. On the same chip, Framework’s 192 GB Gorgon Halo box wrote DeepSeek V4.1 Flash at 16 tok/s on a 32K prompt with drafting, a model too large for any 128 GB Spark.
Gorgon Halo plus a graphics card: where the Lucebox fits
We make the Lucebox Zero 495, so read this box as ours.
Where it fits
- The Lucebox is a Gorgon Halo box with an R9700 card added. On DeepSeek V4.1 Flash at a 32K prompt it wrote 33 tok/s and read 1,000 in our lab; Framework published 16 and 340 for the same chip without the card, on its own engine and file.
- Qwen3.8-27B runs on the card instead of the chip: the R9700 wrote it 2.9x faster than a Strix Halo box on the same hipfire build, file and prompt (2.6x without drafting).
Where it doesn't
- Your tools need CUDA, which neither the Lucebox nor any Strix Halo box runs.
- A plain Strix Halo box costs less: from $3,500 against $5,999.
- You need one in hand 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
- 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.
Apple Mac Studio
DGX Spark vs Strix Halo FAQ
Strix Halo or GB10: which chip is faster?
Strix Halo and GB10 write large models at similar speeds, 256 against 273 GB/s of memory bandwidth, and on DeepSeek V4 Flash with drafting the best runs we have are 41.9 tok/s on Strix Halo (our lab) and 35.3 on GB10 (a published run). The prompt-reading numbers we have favour GB10: 822.98 tok/s at 64K in ds4’s baseline, against 340 for Gorgon Halo on the larger DeepSeek V4.1 Flash, a different model.
Is a Ryzen AI MAX+ 395 box good for local AI models?
A Ryzen AI MAX+ 395 box is good for models up to about 100 GB: its 128 GB holds DeepSeek V4 Flash (101.5 GB) and Qwen3.8-Flash-Next (106.3 GB), models no single graphics card holds. With drafting it wrote them at 41.9 and 51.39 tok/s. On a 27B model a 32 GB card is 2.9x faster (2.6x without drafting).
Is a Strix Halo mini PC good enough for local LLMs?
For one person running open models without CUDA, a Strix Halo mini PC is good enough: $3,500 buys 128 GB, and the big MoE models write at 30.75 to 51.39 tok/s in the runs above. It will not hold DeepSeek V4.1 Flash, and for a Mac at the same budget see Strix Halo vs Mac Studio.
What is Gorgon Halo?
Gorgon Halo is the AMD Ryzen AI MAX+ PRO 495, the step up from Strix Halo: 16 Zen 5 cores, a 40-CU Radeon 8065S and up to 192 GB of LPDDR5X-8533 at 273 GB/s, per guru3d. The Framework Desktop with it costs $6,799 and is sold out until its third batch in December.