Llama-3.1-Nemotron-Nano-VL-8B on DGX Spark — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Llama-3.1-Nemotron-Nano-VL-8B
Parameters	8.7B
Device	DGX Spark
Quantization	BF16
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	8
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	0.89	14.2	≤3 s
2	1.04	13.6	≤3 s
4	1.44	12.0	≤3 s
8	2.17	9.6	≤3 s
16	4.43	7.2	>3 s
32	8.30	4.2	>3 s
64	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.89	14.2	≤3 s
2	1.73	14.0	≤3 s
4	3.58	13.6	>3 s
8	7.80	12.6	>3 s
16	18.3	11.3	>3 s
32	47.6	9.2	>3 s
64	—	—	Failed

Llama-3.1-Nemotron-Nano-VL-8B (8.7B parameters), served in BF16 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.89 s, then writes 14.2 tok/s.

With 32 images per camera at 720p, the answer starts after 47.61 s. With one 2K image, after 0.92 s.

At the default target of an answer starting within 3 s (720p, one image per camera), it keeps up with 8 cameras at once.