Sunk Cost sunkcost.ai Data checked 2026-09-03

Can a DGX Spark GB10 Grace Blackwell, 128GB run local LLMs?

Yes — 33 of the 39 open models on this site fit in its 119.5 GB of usable memory, the strongest being Qwen3.8 27B. Whether that saves you money is a different question, and the answer is usually no.

Price$4,699
Memory128 GB, about 119.5 GB of it addressable by the GPU at 273 GB/s
Best model it runsQwen3.8 27B — Sonnet-class, 11 tok/s
Pay-back against the APIPays back in 62 years at 500k tokens a day

Run the numbers on this machine

What it runs

ModelSpeedClassGood atMemory
Qwen3.8 27BQ4_K_M 11 tok/s Sonnet-class 19 GB
Ling 3.0 flashQ4_K_M 26 tok/s Haiku-class 82 GB
Qwen3.6 27BQ4_K_M 11 tok/s Haiku-class 19 GB
Qwen3.6 35B-A3BQ4_K_M 70 tok/s Haiku-class 23 GB
Muse Glimmer 30BQ4_K_M 11 tok/s Haiku-class 18 GB
Gemma 4 26B-A4BQ4_K_M 52 tok/s Haiku-class 18 GB
Qwen3.5 122B-A10BUD-Q4_K_M 25 tok/s Haiku-class 79 GB
Gemma 4 31B itQ4_K_M 7.9 tok/s Haiku-class 26 GB
Granite 4.2 30BQ4_K_M 7.8 tok/s Haiku-class 26 GB
GLM-4.7-FlashQ4_K_M 50 tok/s Haiku-class 20 GB
Gemma 4 12BQ4_K_M 26 tok/s Below every hosted tier 8.0 GB
Qwen3.5 9BQ4_K_M 30 tok/s Below every hosted tier 6.8 GB

21 more fit; the calculator lists them all.

The specifics

Chip
GB10 Grace Blackwell — Founders Edition, 4TB
Memory bandwidth
273 GB/s
Usable by the GPU
119.5 GB — CUDA deviceQuery reports 122,570 MiB (about 119.7 GiB) visible to the GPU under DGX OS; treated as 119.5 GB usable. Partner boxes (ASUS, Dell, Lenovo, Acer) exist but none was verifiably cheaper and in stock when checked.
Power under load
150 W (third party measured) — NVIDIA publishes no system draw (240 W supply, GB10 TDP 140 W). 150 W is a single third-party unit's peak under inference (kubesimplify); Jeff Geerling measured 122.8 W running gpt-oss-20b.
Availability
Shipping. MSRP raised from $3,999 to $4,699 in February 2026.
Sources
source 1, source 2, source 3, source 4