Nemotron-Nano-12B-v2-VL on DGX Spark — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Nemotron-Nano-12B-v2-VL
Parameters	13B
Device	DGX Spark
Quantization	BF16
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	16
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	0.57	10.0	≤3 s
2	0.78	9.5	≤3 s
4	0.88	9.2	≤3 s
8	1.37	7.7	≤3 s
16	2.67	6.3	≤3 s
32	4.72	4.7	>3 s
64	8.58	3.0	>3 s
72	9.97	2.8	>3 s
80	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.57	10.0	≤3 s
2	0.93	10.0	≤3 s
4	1.82	10.9	≤3 s
8	3.60	12.0	>3 s
16	6.78	10.3	>3 s
32	13.6	9.8	>3 s
64	27.5	9.3	>3 s
128	59.0	9.1	>3 s
256	—	—	Failed

Nemotron-Nano-12B-v2-VL (13B parameters), served in BF16 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.57 s, then writes 10 tok/s.

With 128 images per camera at 720p, the answer starts after 59.03 s. With one 2K image, after 1.36 s.

At the default target of an answer starting within 3 s (720p, one image per camera), it keeps up with 16 cameras at once.