Cosmos3-Nano on DGX Spark — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Cosmos3-Nano
Parameters	16B
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	0.47	14.2	≤3 s
2	0.50	14.5	≤3 s
4	0.57	13.7	≤3 s
8	0.86	12.0	≤3 s
16	2.00	10.8	≤3 s
32	3.18	7.9	>3 s
64	5.71	4.6	>3 s
96	8.62	3.2	>3 s
104	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.47	14.2	≤3 s
2	0.65	14.1	≤3 s
4	1.28	13.9	≤3 s
8	2.60	13.6	≤3 s
16	5.53	13.1	>3 s
32	12.8	11.8	>3 s
64	33.7	9.6	>3 s
128	96.9	7.0	>3 s
256	308.9	4.7	>3 s
512	—	—	Failed

Cosmos3-Nano (16B parameters), served in BF16 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.47 s, then writes 14.2 tok/s.

With 256 images per camera at 720p, the answer starts after 308.95 s. With one 2K image, after 1.61 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.