Head to head · NVIDIA vs NVIDIA

DGX Spark vs RTX 5090 for local LLMs

They solve different problems. The RTX 5090 has 32 GB at 1,792 GB/s: about 6.6× the Spark's bandwidth, so models that fit run roughly 3.2–6.6× faster (our estimate). On the one model both have public runs for, gpt-oss-20b, the Spark measured 83 tok/s and the 5090 282 tok/s. The DGX Spark has 128 GB (~109 GB usable) at 273 GB/s, so it holds 6 models we track that a single 5090 can't, up to Qwen3.5 122B-A10B. Prices & stock checked 2026-10-24. Prices change daily and some retailers block our checker, so always confirm on the store page.

Specs that matter

	DGX Spark / GB10	RTX 5090
Price	$5,859cheapest GB10 box, complete system	$8,500card only, needs a host PC
Memory	128 GB unified (~109 GB usable)	32 GB VRAM (~30 GB usable)
Bandwidth	273 GB/s	1,792 GB/s
Models that fit (4-bit)	14 of 16	8 of 16

Model by model (4-bit)

Model	DGX Spark	RTX 5090
Qwen3.5 9B~7 GB at 4-bit	Fits~39 tok/s	Fits~214 tok/s
Gemma 4 12B~9 GB at 4-bit	Fits~29 tok/s	Fits~170 tok/s
GPT-OSS 20B~13 GB at 4-bit	Fits~80 tok/s	Fits~264 tok/s
Gemma 4 26B-A4B~18 GB at 4-bit	Fits~70 tok/s	Fits~247 tok/s
Qwen3.8 27B~19 GB at 4-bit	Fits~13 tok/s	Fits~83 tok/s
GLM-4.7-Flash 30B-A3B~21 GB at 4-bit	Fits~83 tok/s	Fits~269 tok/s
Gemma 4 31B~21 GB at 4-bit	Fits~11 tok/s	Fits~73 tok/s
Qwen3.6 35B-A3B~24 GB at 4-bit	Fits~83 tok/s	Fits~269 tok/s
Llama 3.3 70B~46 GB at 4-bit	Fits~5 tok/s	Won't fit
Qwen3-Next 80B-A3B~52 GB at 4-bit	Fits~83 tok/s	Won't fit
Llama 4 Scout 109B-A17B~71 GB at 4-bit	Fits~19 tok/s	Won't fit
GPT-OSS 120B~67 GB at 4-bit	Fits~61 tok/s	Won't fit
Nemotron 3 Super 120B-A12B~78 GB at 4-bit	Fits~26 tok/s	Won't fit
Qwen3.5 122B-A10B~79 GB at 4-bit	Fits~31 tok/s	Won't fit

Estimated single-user decode at ~8K context; 5090 figures assume the model is fully in VRAM. How we estimate

Measured runs vs our formula

Hardware	Model / run	Measured tok/s	Our formula	Source
DGX Spark273 GB/s	Qwen2.5-Coder 7B Q8_0llama.cpp tg32	29.4	26.0	source
DGX Spark273 GB/s	Gemma 3 4B Q4_0llama.cpp tg32	81.0	83.4	source
DGX Spark273 GB/s	gpt-oss-120b MXFP4llama.cpp tg32	58.7	61.4	source
DGX Spark273 GB/s	gpt-oss-20b MXFP4llama.cpp tg32	83.4	80.0	source
DGX Spark273 GB/s	Qwen3-Coder 30B-A3B Q8_0llama.cpp tg32	61.1	49.8	source
DGX Spark273 GB/s	GLM-4.7-Flash Q8_0llama.cpp tg32	48.7	53.9	source
DGX Spark273 GB/s	Qwen3.5-122B-A10B Q5_K_Sllama.cpp tg128	23.0	27.7	source
Dell Pro Max GB10273 GB/s	Llama 3.1 70B Q4_K_MOllama (Geerling)	4.7	4.9	source
RTX 50901,792 GB/s	Llama-2 7B Q4_0llama.cpp tg128	290.0	280.4	source
RTX 50901,792 GB/s	Qwen3.6-27B Q4_K_Mllama.cpp tg128	77.1	81.3	source
RTX 50901,792 GB/s	Qwen3.6-35B-A3B UD-Q4_K_Mllama.cpp tg128	271.0	268.6	source
RTX 50901,792 GB/s	gpt-oss-20b MXFP4llama.cpp tg128	282.5	264.1	source
RTX 50901,792 GB/s	Qwen3 30B-A3B Q4_K_Mllama.cpp + GGML_CUDA_GRAPH_OPT=1	352.1	260.1	source
RTX 50901,792 GB/s	gpt-oss-20b MXFP4llama.cpp + GGML_CUDA_GRAPH_OPT=1	419.1	264.1	source

Public single-user decode results we calibrate against. Runs differ in build and settings, so compare like with like.

Prices

DGX SparkOut of stock

Lenovo ThinkStation PGX

1TB

$5,859at B&H Photo

128 GB · 273 GB/s

Check price

RTX 5090Out of stock

NVIDIA GeForce RTX 5090 Founders Edition

32GB card

$8,500at Newegg

32 GB VRAM · 1,792 GB/s

Card only: you still need a host PC.

Check price

FAQ

Is DGX Spark faster than an RTX 5090?

No. On models that fit in 32 GB, the 5090 is about 3.2–6.6× faster by our estimate, because it has 1,792 GB/s versus 273 GB/s. On the one model both have public runs for, gpt-oss-20b, the Spark measured 83 tok/s and the 5090 282 tok/s.

Why buy a DGX Spark instead of a 5090?

Memory. The Spark's ~109 GB usable holds models like GPT-OSS 120B, Nemotron 3 Super 120B-A12B, Qwen3.5 122B-A10B that won't fit in a single 5090's ~30 GB, and it's a complete computer.

What do they cost?

The cheapest GB10 box we track is $5,859 (Lenovo ThinkStation PGX), a complete system. The RTX 5090 Founders Edition is $8,500 for the card only; you still need a host PC, which we don't price.

Can the RTX 5090 run Llama 70B?

Not fully in VRAM: it needs ~46 GB at 4-bit. The Spark fits it but runs it at ~5 tok/s.

Which is better for coding models like Qwen3.6 35B-A3B?

If it fits, the 5090: we estimate ~269 tok/s vs ~83 on the Spark at 4-bit.