$6,950 for 128 GB
The DGX Spark launched at $3,999 and is reported at $6,950 to $6,995 now. Dell sells the same chip for $6,593, and the 64 GB version ($4,999) comes only from other brands.
NVIDIA DGX Spark review, measured
The NVIDIA DGX Spark is a 150 x 150 mm desktop with NVIDIA’s GB10 chip, 128 GB of unified memory and CUDA. Below: its price today, what fits in 64 and 128 GB, and how fast it runs the models a Lucebox runs, next to the machines people weigh it against.
Slow to write, quick to read
On the same engine and file, an M5 Max writes DeepSeek V4 Flash 2.2x faster, and the Spark reads a 64K-token prompt 2.1x faster.
CUDA in a small box
The DGX Spark runs NVIDIA’s full CUDA software in 150 x 150 mm and links to a second Spark over its 200 Gbps port.
DGX Spark specs: what is inside?
Inside the DGX Spark is NVIDIA’s GB10 chip with 128 GB of LPDDR5X at 273 GB/s, a ConnectX-7 network card at 200 Gbps, 10 GbE and Wi-Fi 7, in a box that weighs 1.2 kg.
| Spec | DGX Spark (GB10) |
|---|---|
| Memory | 128 GB LPDDR5X, or 64 GB from other brands |
| Memory speed | 273 GB/s, 256-bit bus |
| Processor and GPU | 20 Arm cores and a Blackwell GPU, 1 PFLOP at FP4 |
| Software | CUDA, DGX OS |
| Power | 140 W chip, 240 W adapter; 104 to 160 W at the wall under load |
| Networking | ConnectX-7, 200 Gbps, links two Sparks |
| Size | 150 x 150 x 50.5 mm |
| Sources | NVIDIA, ServeTheHome |
DGX Spark memory bandwidth
The DGX Spark reads its 128 GB of LPDDR5X at 273 GB/s over a 256-bit bus. Each token a model writes means reading its active weights from memory, so bandwidth sets the ceiling on writing speed: a Mac Studio M5 Max reads at 614 GB/s and an RTX 5090 at 1,792 GB/s.
The GB10 chip in the DGX Spark
The GB10 is one chip, co-designed by NVIDIA and MediaTek, with 20 Arm CPU cores (10 Cortex-X925 and 10 Cortex-A725) and a Blackwell GPU with fifth-generation Tensor Cores, joined by NVLink-C2C so both work from the same memory. NVIDIA rates it at up to 1 petaflop of FP4 compute with sparsity. Acer’s and MSI’s launch coverage puts the GPU at 6,144 CUDA cores, a figure NVIDIA does not list.
DGX Spark power consumption
The DGX Spark runs from a 240 W power adapter, and NVIDIA rates the GB10 chip at 140 W. Under load it draws 104 to 160 W at the wall in ServeTheHome’s tests, which also found GB10 boxes from Gigabyte and HP idling at 36 to 40 W with the network card up and peaking just under 200 W.
DGX Spark memory (VRAM): 64 GB or 128 GB?
The DGX Spark has 128 GB of LPDDR5X shared by its CPU and GPU, so the same 128 GB serves as its VRAM. The 64 GB GB10, sold only by partner brands from $4,999, holds Qwen3.8-27B but not DeepSeek V4 Flash or Qwen3.8-Flash-Next.
| Model and size | GB10 box with 64 GB (Dell, ASUS and others)64 GB unified | NVIDIA DGX Spark128 GB unified |
|---|---|---|
| Qwen3.8-27B15.2 GB | Yes | Yes |
| DeepSeek V4 Flash101.5 GB | No | Yes |
| Qwen3.8-Flash-Next106.3 GB | No | Yes |
| Qwen3.8-27B and DeepSeek V4 Flash at once116.7 GB | No | Just, little room left |
| DeepSeek V4.1 Flash190.8 GB | No | No |
DGX Spark benchmarks: how many tokens per second?
The DGX Spark writes DeepSeek V4 Flash at 18.05 tokens per second after a 2K-token prompt and 13.84 after a 64K one, against 39.35 and 27.64 on a Mac Studio M5 Max, in antirez’s matched ds4 sweep with no drafting.
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.
These runs differ in engine, file or prompt. Drafting speeds some prompts far more than others.
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.
Lucebox (395 build) answers 2.1x as fast as DGX Spark for one user and 2.4x for 8 users at once, same files, Lucebox engine on the Lucebox (395 build) and llama.cpp on the DGX Spark.
DGX Spark answers 1.5x as fast as Lucebox (395 build) for 4 users at once, each machine's own tuned setup.
DGX Spark with vLLM, llama.cpp and Ollama
NVIDIA publishes DGX Spark playbooks for vLLM (also across several linked Sparks), llama.cpp, SGLang, NIM and LM Studio. Ollama has no playbook, but it was a launch partner and says it runs on the Spark out of the box. Inference speed depends on the engine as much as on the chip: the sweep above used DwarfStar (ds4), and our own image-question runs on the Spark used llama.cpp.
DGX Spark speed test: watch it write
The same answer at the measured writing speed, on runs made the same way or each machine’s best setup, as the line says.
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
DGX Spark price: what does it cost now?
The DGX Spark Founders Edition sells for $6,950 to $6,995 in the US with 128 GB of memory and 4 TB of storage, per the press reports we last checked on 30 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 |
US prices read on 30 October 2026.
- $3,999 from 13 October 2025. Launch price. Yahoo Tech
- $4,699 from 23 February 2026. NVIDIA raises it, citing memory supply. NVIDIA forum
- $6,950 from 2 October 2026. Raised again as a 64 GB version arrives. The Register
Founders Edition or a partner version?
NVIDIA’s own DGX Spark, the Founders Edition, comes in a single configuration. Acer, ASUS, Dell, Gigabyte, HP, Lenovo and MSI sell GB10 boxes built to NVIDIA’s design: reviews find the same chip and memory, and differences in chassis, cooling, storage, price and support. Partner boxes with 128 GB start at $5,859, and each one is set against the Spark on DGX Spark alternatives.
What else can you buy for the price of a DGX Spark?
Set a budget and pick a model to see which machines run it, with the DGX Spark as one of the rows.
Model
Rank by
Our pick
Lucebox Zero 495, 128 GB
- Holds DeepSeek V4 Flash (101.5 GB) across the 32 GB card and 128 GB of unified memory.
- Writes it at 51.1 tok/s on our earlier 395 build, drafting on (our lab).
- $5,999, $1,001 under your budget.
Also consider
- Holds DeepSeek V4 Flash (101.5 GB) in 128 GB of unified memory.
- Writes it at 41.9 tok/s on another box with the same chip, drafting on (our lab).
- $3,500, $3,500 under your budget.
We make the Lucebox. It is ranked by the same rule as every other row; a run marked “on our earlier 395 build” comes from the previous Lucebox, with the same R9700 card.
- Lucebox Zero 495, 128 GB32 GB card + 128 GBCard + unified 63% used273 GB/s51.1 tok/s drafting on, our lab, on our earlier 395 build
- GMKtec EVO-X2, 128 GB128 GB unifiedFits 79% used256 GB/s41.9 tok/s drafting on, our lab, on another box with the same chip
- NVIDIA DGX Spark128 GB unifiedFits 79% used273 GB/s35.3 tok/s drafting on, public benchmark
- Mac Studio (M5 Max), 128 GB128 GB unifiedFits 79% used614 GB/s39.35 tok/s no drafting, public benchmark
- Framework Desktop (Ryzen AI MAX+ PRO 495)192 GB unifiedFits 53% used273 GB/sno run yet
- GMKtec EVO-X2, 64 GB64 GB unifiedDoes not fit needs 159%256 GB/sno run yet
- Mac Studio (M5 Max), 48 GB48 GB unifiedDoes not fit needs 211%614 GB/sno run yet
- GB10 box with 64 GB (Dell, ASUS and others)64 GB unifiedDoes not fit needs 159%273 GB/sno run yet
- Mac Studio (M5 Ultra), 96 GB96 GB unifiedDoes not fit needs 106%1,228 GB/sno run yet
- RTX 5090 card32 GB on the cardDoes not fit needs 317%1,792 GB/sno run yet
- Lucebox Zero 495, 192 GB32 GB card + 192 GB$1,199 over budgetCard + unified 45% used273 GB/s51.1 tok/s drafting on, our lab, on our earlier 395 build
- Mac Studio (M5 Ultra), 256 GB256 GB unified$3,799 over budgetFits 40% used1,228 GB/sno run yet
- RTX PRO 6000 workstation96 GB on the cardDoes not fit needs 106%1,792 GB/sno run yet
- tinybox red flash192 GB on 6 cards$15,000 over budgetOn the card 53% used3,840 GB/sno run yet
- Mac Studio (M5 Ultra), 512 GB512 GB unified$15,299 over budgetFits 20% used1,228 GB/sno run yet
The pick holds the model with room for the prompt, fits your budget and has the fastest run we can link to. Runs with drafting rank against each other first, then runs without it, then machines with no run, by memory speed. The amber band is the Lucebox's 32 GB card.
Is the DGX Spark worth it? Who should buy one
The DGX Spark is worth its price if your software needs CUDA or your work is mostly long prompts, and a poor buy for chat or for models of 32B and under.
Buy one if
- Your software needs CUDA, or you build for NVIDIA’s bigger DGX systems.
- Your work is long prompts: on the same engine and file it reads a 64K-token prompt 2.1x faster than an M5 Max.
- You do fine-tuning: NVIDIA says the 128 GB Spark can fine-tune models up to 70B parameters and publishes guides for NeMo, Unsloth and LLaMA-Factory.
- You want two of them linked over their 200 Gbps ports later.
Look elsewhere if
- You mostly chat or write: an M5 Max writes 2.2x faster on the same engine and file, for $5,399, as the DGX Spark vs Mac Studio runs show.
- Your models are 32B or smaller: an RTX 5090 reads memory 6.6 times faster, which DGX Spark vs RTX 5090 covers.
- You want DeepSeek V4.1 Flash on one box: its 2-bit weights, about 191 GB, are larger than the Spark’s memory.
DGX Spark vs Lucebox: where each fits
The Lucebox Zero 495 is a workstation with a Ryzen AI MAX+ PRO 495 and a Radeon AI PRO R9700 card. Our earlier 395 build, with the same card, wrote DeepSeek V4 Flash at 51.1 tokens per second with DSpark drafting on our engine and file (our run). One Spark wrote 35.3 with DSpark on ds4 and antirez’s Q2 file in an independent run, on a shorter prompt.
We make the Lucebox Zero 495, so read this box as ours.
Where it fits
- Image questions on Qwen3.8-27B: eight at once in 10.5 s on the same R9700 in our earlier 395 build, against 25.4 s on a DGX Spark with the same files, our engine against llama.cpp on the Spark.
- DeepSeek V4.1 Flash on one box: 33 tok/s at a 32K prompt with 192 GB in our lab. The 128 GB model runs it too, reading rarely used experts from the SSD.
- DeepSeek V4 Flash with drafting: 51.1 tok/s on our 395 build, against 35.3 published for one Spark, each with its own setup.
Where it doesn't
- Your software needs CUDA.
- DeepSeek V4 Flash with images: the Spark answered four at once in 24.6 s, we took 37.9 s.
- You want the smallest box: the Spark is far smaller.
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.
Apple Mac Studio
DGX Spark FAQ
What is the DGX Spark?
The NVIDIA DGX Spark is a small desktop machine for running and fine-tuning AI models, built on the GB10 chip with 20 Arm CPU cores, a Blackwell GPU and 128 GB of unified memory. It runs NVIDIA’s DGX OS with CUDA and sells for $6,950 to $6,995 in the US.
How much does the NVIDIA DGX Spark cost?
The NVIDIA DGX Spark costs $6,950 to $6,995 in the US as of October 2026, up from $3,999 at launch and $4,699 after February 2026. Other brands sell the same GB10 chip with 128 GB from $5,859, and with 64 GB from $4,999.
Why did the DGX Spark price go up?
NVIDIA raised the DGX Spark from $3,999 to $4,699 on 23 February 2026 “due to memory supply constraints”, in its own announcement. The rise to $6,950 in October 2026 came with no NVIDIA statement we could find; The Register and ServeTheHome tie it to memory prices. Partner boxes moved too: Acer’s GB10 box launched at $3,999, and 128 GB GB10 boxes from other brands now start at $5,859.
When did the DGX Spark come out?
NVIDIA announced on 13 October 2025 that the DGX Spark was shipping, and orders opened on NVIDIA.com and at Micro Center in the US on 15 October 2025. It launched at $3,999 with 128 GB.
Is there a 64 GB DGX Spark?
Since 23 October 2026, but only from other brands such as Dell, ASUS, Acer, Gigabyte, HP and MSI, from $4,999. NVIDIA’s own Spark has 128 GB.
Which models can the DGX Spark run?
The 128 GB DGX Spark holds Qwen3.8-27B (15.2 GB), DeepSeek V4 Flash (101.5 GB) and Qwen3.8-Flash-Next (106.3 GB), and just fits Qwen3.8-27B and DeepSeek V4 Flash together (116.7 GB). DeepSeek V4.1 Flash, about 191 GB at 2 bits, does not fit, and a 64 GB GB10 box holds only Qwen3.8-27B of these.
Can the DGX Spark run a 70B model?
Yes. A 70B model at 4-bit is about 40 GB, which fits in the DGX Spark’s 128 GB with plenty left for a long context.
Can one DGX Spark run DeepSeek V4.1 Flash?
No: at 2 bits DeepSeek V4.1 Flash needs about 191 GB, more than the DGX Spark’s 128 GB, and DeepSeek V4.1 Flash on the Lucebox shows a 128 GB workstation running it with its rarely used experts on the SSD.
Can you cluster two DGX Sparks?
Yes. Each DGX Spark has a ConnectX-7 network card with two 200 Gbps QSFP ports, and NVIDIA supports up to four linked Sparks: two connect with one QSFP cable, and four need an external 200 GbE switch. NVIDIA publishes guides for pairing Sparks and for running vLLM across them.
Does the DGX Spark run Windows?
No. The DGX Spark ships with DGX OS (Ubuntu Linux), and NVIDIA offers no Windows for it; its Windows machines on GB10-class chips are a separate line called RTX Spark. StorageReview found a Windows boot certificate in a Spark firmware update in September 2026, but NVIDIA has announced no Windows drivers or installer.
Can you game on the DGX Spark?
Some games run, but the DGX Spark is not built for gaming: it runs Linux on Arm cores, so Steam games go through x86 translation and Proton. One owner, an NVIDIA employee who bought his own unit, reports Cyberpunk 2077 and Indiana Jones running at about two-thirds of an RTX 5070. Games whose anti-cheat blocks Linux, such as League of Legends, do not run.