DGX Spark 128G — what LLMs can it run?

The short answer for DGX Spark 128G

128 GB, 273 GB/s Unified LPDDR5X (One pool shared by the GB10's CPU and GPU and by DGX OS itself. NVIDIA publishes no fixed GPU share, so fits are judged against the whole pool — and nothing larger than it can load, because there is no separate system RAM to spill into.). 102 of 112 models in this index fit comfortably at 4K context; 103 load at all.

Biggest that fits

DBRX Instruct — 84.3 GB at Q4_K_M — 43.7 GB spare at 4K, so longer context comes out of a thin margin.

Room to grow

Nemotron 3.5 Super VL 120B-A12B — 75.1 GB at Q4_0 — under 60% of the card, which leaves 52.9 GB for a long context window or a second process.

Speed ceiling

Llama 3.1 8B Instruct at Q4_K_M reads 4.6 GB of weights per token, so 273 GB/s puts a hard ceiling near 59 tok/s. That is arithmetic on two published numbers, not a benchmark — real throughput lands below it. No run on this card has been measured here, so there is nothing to compare the ceiling against.

The wall you will hit

Bandwidth, and one pool. 128 GB holds models no single consumer card can, but every byte of it — CPU, GPU and DGX OS alike — comes from the same LPDDR5X at 273 GB/s, a fraction of a discrete flagship's. There is no system RAM behind it to spill into, so this site treats anything larger than the pool as not loading at all. The largest dense model here, Llama 3.2 90B Vision Instruct at Q4_K_M, has a ceiling of about 5 tok/s on it; mixture-of-experts models read only their active experts per token and run far faster than their size suggests.