NVIDIA DGX Spark 128 GB
A GB10 Grace Blackwell desktop appliance with 128 GB of coherent unified LPDDR5X. Its 273 GB/s of bandwidth is modest, so it is far better at prefill and at sparse MoE models than at dense models. Its real selling point is a full CUDA software stack in a 170 W box.
NVIDIAapplianceAvailable23 models fit3 measured
Memory
128 GB
Official spec LPDDR5X unified (coherent)
Memory bandwidth
273 GB/s
Official spec ~67% achieved in practice
Specifications
CPU
NVIDIA GB10 Grace, 20-core Arm Official spec
GPU
Blackwell GB10 GPU Official spec
Memory
128 GB LPDDR5X unified (coherent) Official spec
Memory bandwidth
273 GB/s Official spec source
Achieved bandwidth ⓘ
~67% (183 GB/s) Measured
FP16 compute ⓘ
~125 TFLOPS Official spec
Architecture
Unified LPDDR (integrated accelerator) Official spec
Nominal power
240 W Official spec
Measured load power
170 W Measured
Idle power
25 W Measured
Released
15 Oct 2025 Official spec
The bandwidth-to-compute ratio is unusual: prefill is excellent for the size of the box (measured at 1,956 t/s on gpt-oss-120b) while decode is limited by 273 GB/s. Buy it for the CUDA stack and MoE models, not for dense 70B. Achieved-bandwidth coefficient CALIBRATED to 67% from 6 measured runs on this configuration.
Standout models on this machine
Fastest
gpt-oss-20b — ~59.4–85.5 t/s
MXFP4 · estimated confidence
Largest that fits
Compare against
Related hardware
Framework Desktop (Ryzen AI Max+ 395, 128 GB)128 GB · 256 GB/sGMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB)128 GB · 256 GB/sQuad RTX 5090 workstation (4x 32 GB)128 GB · 7,168 GB/sMac Studio M5 Max 128 GB128 GB · 614 GB/sDual used RTX 4090 workstation (48 GB)48 GB · 2,016 GB/sRTX PRO 6000 Blackwell workstation (96 GB)96 GB · 1,792 GB/s
Model performance on NVIDIA DGX Spark 128 GB3 measured, 20 estimated
23 of 23 rows
| Model↕ | Quant | Memory↕ | Decode▼ | Prefill↕ | Context | Fit | Confidence↕ |
|---|---|---|---|---|---|---|---|
| gpt-oss-20b OpenAI · 20.915B (3.6B active) | MXFP4recommended | 13.0 GB | ~59.4–85.5 t/s | ~2880–5980 t/s | 128K | Comfortable | Estimated |
| gpt-oss-120b OpenAI · 116.829B (5.1B active) | MXFP4recommended | 66.5 GB | 56.4 t/s | 1680 t/s | 128K | Comfortable | High |
| Nemotron 3.5 Lightning 30B-A3B NVIDIA · 31.578B (3B active) | Q8_0recommended | 33.9 GB | ~37.3–53.6 t/s | ~1280–2670 t/s | 1024K | Comfortable | Estimated |
| Qwen3-Coder 30B-A3B Alibaba Qwen · 30.532B (3.3B active) | Q8_0recommended | 33.5 GB | ~34.1–49.1 t/s | ~1240–2580 t/s | 256K | Comfortable | Estimated |
| Qwen3 30B-A3B Alibaba Qwen · 30.532B (3.3B active) | Q8_0recommended | 33.5 GB | ~34.1–49.1 t/s | ~1240–2580 t/s | 40K | Comfortable | Estimated |
| Mistral Small 4 119B-A6.5B Mistral AI · 119.401B (6.5B active) | Q4_K_Mrecommended | 72.7 GB | ~30.5–44 t/s | ~448–931 t/s | 256K | Comfortable | Estimated |
| Gemma 4 26B-A4B Google DeepMind · 25.806B (3.8B active) | Q8_0recommended | 29.7 GB | ~29.9–43 t/s | ~1260–2620 t/s | 256K | Comfortable | Estimated |
| Qwen3.5 122B-A10B Alibaba Qwen · 125.086B (10B active) | Q4_K_Mrecommended | 76.1 GB | ~20.3–29.1 t/s | ~353–734 t/s | 256K | Comfortable | Estimated |
| Nemotron 3 Super 120B-A12B NVIDIA · 123.611B (12B active) | Q4_K_Mrecommended | 75.3 GB | ~17–24.4 t/s | ~324–674 t/s | 1024K | Comfortable | Estimated |
| Qwen3 8B Alibaba Qwen · 8.191B | Q8_0recommended | 10.6 GB | ~16.7–24 t/s | ~1530–3170 t/s | 40K | Comfortable | Estimated |
| DeepSeek-V4 Flash DeepSeek · 304.18B (13B active) | DS4 Q2 (mixed)recommended | 85.1 GB | 18.1 t/s | 826 t/s | 192K | Fits | Medium |
| Phi-4 14B Microsoft · 14.66B | Q8_0recommended | 17.8 GB | ~9.5–13.6 t/s | ~852–1770 t/s | 16K | Comfortable | Estimated |
| DeepSeek-V4.1 Flash DeepSeek · 763.205B (16B active) | DS4 V4.1 Q2 (SSD)recommended | 93.2 GB | 9.4 t/s | 35.2 t/s | 1024K | Fits | High |
| Devstral Small 24B Mistral AI · 23.572B | Q8_0recommended | 26.8 GB | ~5.9–8.5 t/s | ~530–1100 t/s | 128K | Comfortable | Estimated |
| Mistral Small 3.2 24B Mistral AI · 24.011B | Q8_0recommended | 27.2 GB | ~5.8–8.4 t/s | ~520–1080 t/s | 128K | Comfortable | Estimated |
| Gemma 3 27B Google DeepMind · 27.432B | Q8_0recommended | 33.4 GB | ~5.1–7.3 t/s | ~455–946 t/s | 128K | Comfortable | Estimated |
| Qwen3.8 27B Alibaba Qwen · 27.781B | Q8_0recommended | 30.4 GB | ~5–7.2 t/s | ~450–934 t/s | 256K | Comfortable | Estimated |
| Qwen3.6 27B Alibaba Qwen · 27.781B | Q8_0recommended | 30.4 GB | ~5–7.2 t/s | ~450–934 t/s | 256K | Comfortable | Estimated |
| Granite 4.2 30B IBM · 29.277B | Q8_0recommended | 33.4 GB | ~4.8–6.9 t/s | ~427–886 t/s | 128K | Comfortable | Estimated |
| Gemma 4 31B Google DeepMind · 31.273B | Q8_0recommended | 41.0 GB | ~4.5–6.4 t/s | ~399–830 t/s | 64K | Comfortable | Estimated |
| Qwen3 32B Alibaba Qwen · 32.762B | Q8_0recommended | 37.1 GB | ~4.3–6.2 t/s | ~381–792 t/s | 40K | Comfortable | Estimated |
| DeepSeek-R1-Distill 32B DeepSeek · 32.764B | Q8_0recommended | 37.1 GB | ~4.3–6.2 t/s | ~381–792 t/s | 128K | Comfortable | Estimated |
| Llama 3.3 70B Meta · 70.554B | Q8_0recommended | 76.8 GB | ~2–2.9 t/s | ~177–368 t/s | 64K | Comfortable | Estimated |
Rows in grey are estimates from our bandwidth model, not measurements — they are shown as a range and never as a precise figure. Use the “measured only” filter to see just the 3 pairings on this machine that a real benchmark backs.
Submit a benchmarkContributions are reviewed before publication.
Ownership economicsUnited States (federal) · C corporation · 8h/day
Monthly economic cost
$52
Calculated after tax
Codex/Claude Code
$100/mo
≈ $100/mo · local is $48 less
Net cash at purchase
$3,873
Calculated VAT not reclaimable
Total over 5 years
$3,091
Calculated after tax, after resale
Cost per USD/1M tokens
$5.61
Calculated 9.2M tokens/month
The $100 comparison uses the Codex Pro 5x / Claude Max 5x plans. A fixed planning conversion is used for USD. This compares monthly spend only: subscriptions have usage limits, local hardware has different capabilities and constraints, and taxes or regional pricing may change the charged amount. Prices checked 7 September 2026.
Purchase-price input: $3,873 in United States (federal), tax/VAT excluded · estimated · NVIDIA / partner resellers (converted catalogue estimate). This is the localized purchase input; the after-tax economic cost is calculated separately below.
Monthly breakdown
| Depreciation 3,408 over 5 years, straight-line to a 465 residual | $56.80 |
| Electricity 9.5 kWh/month at 0.140/kWh | $1.33 |
| Cost of capital 4.0%/yr on 2,169 average capital employed | $7.23 |
| Monthly cost before tax | $65.36 |
| Electricity tax shield Running costs are deductible business expenses | −$0.28 |
| First-year expensing §179 (100.0%) | −$13.55 |
| Monthly economic cost after tax | $51.52 |
Three different numbers, deliberately
Cash cost
$3,873
Money that leaves the bank account on day one, net of reclaimable VAT.
Accounting depreciation
$56.80/month
$3,408 written down over 5 years to a $465 residual.
After-tax economic cost
$51.52/month
Depreciation plus running costs plus cost of capital, less the tax those deductions save. This is the figure to compare between machines.
Assumptions and sources (verified 2026-09-07)
VAT rate
0% Official spec
VAT recoverable
0% Official spec
Federal C-corporation estimate at the flat 21% rate. Pass-through entities should use the sole-proprietor estimate as a rougher proxy.
Effective deduction rate
21.00% Calculated
Headline marginal rate 21.00%.
Depreciation
5 years, straight-line Assumption
Residual value
$465 (12%) Assumption
Extrapolated.
Electricity
$0.140/kWh Assumption
Average power draw
54 W Calculated
Load 170 W for 20% of powered hours, idle 25 W for the rest — a machine that is on is not generating tokens the whole time.
Investment allowances
§179 100.0% Official spec
Worth $813 in total — first-year expensing that replaces later tax depreciation.
Cost of capital
$7.23/month Assumption
- VAT rate: No US federal VAT · verified 2026-09-07
- Marginal tax rate: IRS Publication 542 — Corporations · verified 2026-09-07
- VAT recoverable fraction: site assumption · verified 2026-09-07
- Useful life: IRS Publication 946 — How To Depreciate Property · verified 2026-09-07
- Electricity price: Site assumption · verified 2026-09-07
Assumes the machine generates tokens 20% of its 176 powered hours per month. Cost per token scales inversely with this number — halve the utilisation and the cost per token doubles.
Federal planning estimate only, not tax advice. State and local income tax, sales/use tax and incentives are excluded. It assumes 100% business use, a Section 179 election, enough business income to use it, and that the full annual limit remains available.