Hardware › DGX Spark local LLM benchmarks: 128 measured configs

DGX Spark local LLM benchmarks: 128 measured configs

128 measured local-LLM configs on NVIDIA DGX Spark (GB10); fastest 14B+ model Qwen3.5-35B-A3B at 131 tok/s. Official specs, Singapore prices read daily, every number linked to its source.

Snapshot 2026-10-16 · 394 configs on AI boxes and 423 on GPU cards · 123 models. Every number is generated from the tracker snapshot and linked to where it was measured; nothing here is typed by hand.

Silicon: NVIDIA GB10 Grace Blackwell Superchip. Official specs.

Quick answers

What is the fastest model on DGX Spark?: Qwen3.5-35B-A3B at 131 tok/s single-stream (NVFP4, atlas (speculative decode, mtp-2)), source last checked 2026-09-23. Models under 14B parameters are left out of this answer. source
What is the fastest 100B+ model on DGX Spark?: Qwen3.8-Flash-Next at 44.2 tok/s single-stream (NVFP4, vllm (speculative decode, mtp-2)), source last checked 2026-09-01. source
What is the largest model measured on DGX Spark?: Nex-N2-Pro-397B-A17B (397B parameters) at 18.9 tok/s single-stream (llama.cpp), source last checked 2026-10-16. source
How many configs does TokenMark track for DGX Spark?: 128 single-stream configs across 44 models, each linked to the README or model card it came from.

Official specs

gpu: NVIDIA Blackwell architecture GPU, 6,144 CUDA cores, 5th-generation Tensor Cores (FP4 support), 4th-generation RT Cores; 1x NVENC, 1x NVDEC (source)
cpu: 20-core NVIDIA Grace CPU: 10x Arm Cortex-X925 + 10x Cortex-A725 (Armv9) (source)
ai tflops: Up to 1 PFLOP (1,000 TFLOPS / 1,000 TOPS) FP4 (source)
memory gb: 128 (source)
memory type: LPDDR5X, coherent unified system memory (CPU+GPU), 256-bit interface, 16 channels, LPDDR5X-8533 (source)
memory bandwidth gbps: 273 (source)
nvlink c2c: NVLink-C2C coherent CPU<->GPU interconnect present (chip-to-chip within GB10) (source)
networking: NVIDIA ConnectX-7 SmartNIC @ 200 Gbps (2x QSFP connectors, for linking two units); 1x RJ-45 10 GbE; Wi-Fi 7; Bluetooth 5.4 (source)
power w: GB10 TDP 140 W; DGX Spark system power supply 240 W (USB-C PD) (source)
release date: 2025-10-15 (general availability of DGX Spark + partner GB10 systems) (source)

Where to buy NVIDIA DGX Spark (GB10) in Singapore

Read 17 Oct 2026 (SGT) from each seller's own product page. Prices in SGD. Overseas stores are converted at USD/SGD 1.2792 with 9% GST at import added; shipping is not included. Singapore sellers already include GST. List prices on the day, not a quote. No affiliate links.

NVIDIA GB10 box, 128 GB (DGX Spark or ASUS Ascent GX10)

SourceIT (Singapore), ASUS Ascent GX10 1 TB (Authorised reseller, local warranty, GeBIZ): S$6,624
SourceIT (Singapore), DGX Spark 4 TB (Authorised reseller, local warranty, GeBIZ): S$8,096
Vii PC Trade (Singapore), DGX Spark 4 TB (Local reseller, 1yr warranty): S$8,999

Measured single-stream configs on NVIDIA DGX Spark (GB10)

Model	Quant	Backend	Decode	Context	Trust	Source
Qwen3 1.7B	Q4_K_M	llama.cpp	146.1 tok/s	2048	Extracted, source-linked	source
Qwen3.5-35B-A3B	NVFP4	atlas (speculative decode, mtp-2)	131 tok/s		Extracted, source-linked	source
Qwen3.6-35B-heretic	NVFP4	vllm	96 tok/s		Structured table	source
Qwen3.6-35B-A3B-NVFP4-Fast (Unsloth)	NVFP4	vllm	92.6 tok/s		Structured table	source
GPT-OSS-20B	Q4_K_XL	llama.cpp	90.5 tok/s		Structured table	source
Ministral 3B	Q4_K_M	llama.cpp	86.6 tok/s	2048	Extracted, source-linked	source
Qwen3.6-35B-A3B	FP8	metrale (speculative decode, mtp-1)	85 tok/s	2048	Extracted, source-linked	source
Ornith-1.0-35B-AEON	NVFP4	vllm	84.8 tok/s		Structured table	source
Qwen3 30B (MoE)	Q4_K_M	llama.cpp	83.8 tok/s	2048	Extracted, source-linked	source
Qwen3.6-35B-A3B	NVFP4	vllm	77.4 tok/s		Structured table	source
Qwen3.6-35B-A3B	UD-IQ3_XXS	llama.cpp/cuda	66.3 tok/s	8192	Structured table	source
Qwen3.6-35B-A3B	UD-IQ3_S	llama.cpp/cuda	65.6 tok/s	8192	Structured table	source
Qwen3.6-35B-A3B	UD-Q3_K_XL	llama.cpp/cuda	64.1 tok/s	8192	Structured table	source
Qwen3.6-35B-A3B	UD-Q3_K_M	llama.cpp/cuda	64 tok/s	8192	Structured table	source
Qwen3.6-35B-A3B	UD-IQ4_NL	llama.cpp/cuda	63.8 tok/s	8192	Structured table	source
Qwen3.6-35B-A3B	UD-IQ4_XS	llama.cpp/cuda	63.3 tok/s	8192	Structured table	source
Qwen3.6-35B-A3B	UD-IQ4_NL_XL	llama.cpp/cuda	63.1 tok/s	8192	Structured table	source
gemma-4-26B-A4B-it	UD-IQ3_XXS	llama.cpp/cuda	62.9 tok/s	8192	Structured table	source
gemma-4-26B-A4B-it	UD-IQ3_S	llama.cpp/cuda	62.7 tok/s	8192	Structured table	source
Qwen3.6-35B-A3B	Q4_K_XL	llama.cpp	62.7 tok/s		Structured table	source
Qwen3.6-35B-A3B	UD-Q4_K_S	llama.cpp/cuda	62.1 tok/s	8192	Structured table	source
Qwen3.6-35B-A3B	MXFP4	llama.cpp/cuda	61.4 tok/s	8192	Structured table	source
Qwen3.6-35B-A3B	UD-Q4_K_M	llama.cpp/cuda	61 tok/s	8192	Structured table	source
Qwen3.6-35B-A3B	UD-Q3_K_S	llama.cpp/cuda	61 tok/s	8192	Structured table	source
Qwen3.6-35B-A3B	UD-Q4_K_XL	llama.cpp/cuda	59.7 tok/s	8192	Structured table	source

More: all configs · hardware · news · contributors · methodology · llms.txt · llms-full.txt · API (OpenAPI) · JSON snapshot. Built by Altronis, private on-prem AI, Singapore.