DGX Spark 128GB Price & Used Prices: Which Local LLMs Can It Run?
Price alert
We'll email you when the eBay (US) price reaches your target.
NVIDIA's DGX Spark (2025) is a 128GB desktop mini PC running the 176B Qwen3.8 Flash Next at 4-bit at about 43 tokens/s. A 64GB version exists too.
eBay
DGX Spark 128GB price history
eBay
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Full DGX Spark 128GB specifications
- Memory128 GB · LPDDR5X · unified memory
- Bandwidth273 GB/s
- FP16 / BF16125 TFLOPS
- FP8 / INT8250 / 250 TOPS
- Bus width
- 256-bit
- Architecture
- Grace Blackwell (GB10)
- SoC
- NVIDIA GB10
- Cores
- 6,144
- Tensor cores
- 192
- Ecosystem
- CUDA
- NVLink
- Not supported
- Power
- 240 W
- Length
- 150 mm
- Type
- Unified memory
- Launch date
- Oct 15, 2025
- MSRP
- $3,999
Best DGX Spark 128GB deals on eBay
Hand-picked deals from trusted eBay sellers. Shop with confidence.
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
UsedNVIDIA DGX Spark Founders Edition GB10 Grace Blackwell 128GB LPDDR5X 4TB NVMeTop Rated$6,999Buy on eBay
UsedNVIDIA DGX Spark 128GB RAM 4TB SSD - Buy 2 Get Free Connector$7,299+ shippingBuy on eBay
UsedNVIDIA DGX Spark 4TB Founders Edition 940-54242-0000-000 - Open BoxTop Rated$7,999Buy on eBay- NewNVIDIA DGX Spark 4TB Founders Edition 940-54242-0000-000 - New Factory SealedTop Rated$8,999Buy on eBay
New940-54242-0000-000 NVIDIA DGX Spark 20-Core ARM AI 128GB 4TB NVMe Pro Desktop$8,999+ shippingBuy on eBay
NewNVIDIA DGX Spark 4TB Founders Edition GB10 128GB AI Supercomputer SEALED$8,999Buy on eBay
What LLMs can a single DGX Spark 128GB run?
One DGX Spark is best with Qwen3.8 Flash Next at 4-bit; 300B-class MoE needs 2-bit. At 273 GB/s dense models crawl: 31B at 8-bit gets 6 tokens/s.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
 Gemma 4 31B30.7B | 11.8 GB | Runs | 16.3 | 256K |
 Qwen3.6 35B A3B36B (3B active) | 12.3 GB | Runs | 100 | 256K |
| Qwen3.8 Flash Next176B (6B active) | 78.9 GB | Runs | 55.5 | 256K |
 DeepSeek V4 Flash284B (13B active) | 96.8 GB | Runs | 38.5 | 1M |
 Hy3295B (21B active) | 107.2 GB | Runs | 21.1 | 86K |
 GLM 5.3 Flash320B (18B active) | 108.7 GB | Runs | 29.3 | 1M |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 31B30.7B | 18.3 GB | Runs | 10.8 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 22.1 GB | Runs | 71.9 | 256K |
| Qwen3.8 Flash Next176B (6B active) | 111.3 GB | Runs | 42.9 | 256K |
| DeepSeek V4 Flash284B (13B active) | 155.1 GB | Too big | — | — |
| Hy3295B (21B active) | 182.2 GB | Too big | — | — |
| GLM 5.3 Flash320B (18B active) | 199.7 GB | Too big | — | — |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 31B30.7B | 32.6 GB | Runs | 6.2 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 36.9 GB | Runs | 50.5 | 256K |
| Qwen3.8 Flash Next176B (6B active) | 188.2 GB | Too big | — | — |
| DeepSeek V4 Flash284B (13B active) | 161.9 GB | Too big | — | — |
| Hy3295B (21B active) | 317.7 GB | Too big | — | — |
| GLM 5.3 Flash320B (18B active) | 341 GB | Too big | — | — |
RunsFits fully in VRAM; speed is estimated from VRAM bandwidth.
Too bigWon’t fit on one card at this quantization, even with RAM offload.
Full calculation rules are on the Methodology page.
What LLMs can several DGX Spark 128GB units run together?
Two DGX Sparks (256GB) run DeepSeek V4 Flash at 4-bit at about 26 tokens/s; four fit the trillion-parameter Kimi K2.6 at 2-bit, linked at 200Gb/s.
| Model | Size | Units | Total price | Total memory | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 78.9 GB | 1 unit | $8,081 | 128 GB | 51.5 | 240 W |
| DeepSeek V4 Flash284B (13B active) | 96.8 GB | 1 unit | $8,081 | 128 GB | 38 | 240 W |
| Hy3295B (21B active) | 107.2 GB | 1 unit | $8,081 | 128 GB | 15.1 | 240 W |
| GLM 5.3 Flash320B (18B active) | 108.7 GB | 1 unit | $8,081 | 128 GB | 28.8 | 240 W |
| GLM 5.3744B (40B active) | 253.9 GB | 4 units | $32,324 | 512 GB | 13 | 960 W |
| Hy4 preview770B (49B active) | 307.2 GBest. | 4 units | $32,324 | 512 GB | 9.5 | 960 W |
 Muse Spark Open816B (42B active) | 311.2 GB | 4 units | $32,324 | 512 GB | 11.5 | 960 W |
 Kimi K2.7 Code1,000B (32B active) | 339.5 GB | 4 units | $32,324 | 512 GB | 16.1 | 960 W |
| Kimi K2.61,000B (32B active) | 340.5 GB | 4 units | $32,324 | 512 GB | 16.1 | 960 W |
| Model | Size | Units | Total price | Total memory | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 111.3 GB | 1 unit | $8,081 | 128 GB | 40.5 | 240 W |
| DeepSeek V4 Flash284B (13B active) | 155.1 GB | 2 units | $16,162 | 256 GB | 25.6 | 480 W |
| Hy3295B (21B active) | 182.2 GB | 2 units | $16,162 | 256 GB | 10.9 | 480 W |
| GLM 5.3 Flash320B (18B active) | 199.7 GB | 2 units | $16,162 | 256 GB | 16.9 | 480 W |
| GLM 5.3744B (40B active) | 467.3 GB | 4 units | $32,324 | 512 GB | 7.6 | 960 W |
| Hy4 preview770B (49B active) | 467.3 GB | 4 units | $32,324 | 512 GB | 6.5 | 960 W |
| Model | Size | Units | Total price | Total memory | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| DeepSeek V4 Flash284B (13B active) | 161.9 GB | 2 units | $16,162 | 256 GB | 24.7 | 480 W |
| Qwen3.8 Flash Next176B (6B active) | 188.2 GB | 2 units | $16,162 | 256 GB | 26.9 | 480 W |
| Hy3295B (21B active) | 317.7 GB | 4 units | $32,324 | 512 GB | 7.3 | 960 W |
| GLM 5.3 Flash320B (18B active) | 341 GB | 4 units | $32,324 | 512 GB | 10.3 | 960 W |
Full calculation rules are on the Methodology page.
Who it suits
Good fit
- Running 2026's big MoE models on a single machine
Qwen3.8 Flash Next at 4-bit runs at about 43 tokens/s, and at 2-bit DeepSeek V4 Flash does about 38 and GLM 5.3 Flash about 29
- You mainly use MoE models with few active parameters
Qwen3.6 35B A3B at 4-bit runs at about 72 tokens/s, and GLM 4.7 Flash at about 69
- CUDA development and fine-tuning on your desk
NVIDIA says a single unit can fine-tune models up to 70B
- You may want to expand later
Two units connect directly with one QSFP cable for 256GB in total
Not a good fit
- You mainly run dense models
Qwen3.8 27B at 4-bit runs at about 12 tokens/s, only 30% of an RTX 3090's speed
- You depend on x86 or Windows software
With an Arm CPU and DGX OS, many prebuilt packages and the upstream vLLM images don't work as-is
- You want to run DeepSeek V4 Flash or GLM 5.3 Flash at 4-bit on one unit
One won't hold them, so you need two linked together
Before you buy
Make sure it's the 128GB version
On October 2, 2026, NVIDIA also announced a 64GB version: same chip and network ports, half the memory, shipping from Acer, ASUS, Dell, Gigabyte, HP and MSI starting October 23. None of the MoE models of 176B and up fit on the 64GB version, so check the memory capacity when looking at listings.
OEM versions have the same specs
Machines like the ASUS Ascent GX10 and Dell Pro Max GB10 use the same GB10 chip, with exactly the same memory and bandwidth as the DGX Spark, so they run the same models. The only difference is storage capacity, officially up to 4TB.
Software goes through official containers
The OS is DGX OS, based on Ubuntu; the CPU is a 20-core Arm chip; and the GPU uses a new architecture (sm_121). Upstream vLLM images and prebuilt PyTorch packages often don't support this architecture yet, so they install but won't run, and the related issues are still open. NVIDIA's answer is to have users get recent vLLM and PyTorch builds from NGC containers, and its official tutorial site has ready-made recipes for vLLM, SGLang and Ollama. Following those is the least hassle.
nvidia-smi doesn't show memory usage
With unified memory there's no separate VRAM, so nvidia-smi shows Not Supported in the memory usage column; that's normal. The real pitfall is that when frameworks check available memory, they don't count cache the system could free up, so they may wrongly conclude there isn't enough memory. NVIDIA's workaround for now is to manually drop the page cache once and then restart the program.
Use the original power adapter
NVIDIA requires the bundled 240W power adapter; using a different one can cause throttling, failure to boot or unexpected shutdowns. If buying used, confirm the adapter is the original.
Compared with similar machines
DGX Spark 128GB vs Ryzen AI Max+ 395 128GB, Mac Studio M5 Max 128GB, and RTX PRO 6000: price, memory, bandwidth, compute and local LLM speed compared.
| Spec | DGX Spark 128GB | Ryzen AI Max+ 395 128GB | Mac Studio M5 Max 128GB | RTX PRO 6000 |
|---|---|---|---|---|
| Memory | 128 GB | 128 GBSame | 128 GBSame | 96 GB25% less |
| Bandwidth | 273 GB/s | 256 GB/s6% lower | 614 GB/s125% higher | 1,792 GB/s556% higher |
| BF16 / FP16 compute | 125 TFLOPS | 59 TFLOPS52% lower | — | 500 TFLOPS300% higher |
| INT8 compute | 250 TOPS | 59 TOPS76% lower | — | 1,000 TOPS300% higher |
| Qwen3.8 27B speed (2-bit) | 20.2 t/s | 19 t/s6% slower | 43.9 t/s117% faster | 114.6 t/s467% faster |
| Largest dense model (4-bit) | 222B | 222BSame | 222BSame | 165B26% less |
| Power | 240 W | 120 W50% lower | 480 W100% higher | 600 W150% higher |
| Interface | — | — | — | PCIe 5.0 x16 |
| Released | 2025-10 | 2025-03 | 2026-08 | 2025-03 |
| eBay median |
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Amazon.jp
Xianyu