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

Open this configuration in the VLM Inference Benchmark Explorer

Model	Cosmos3-Edge
Parameters	3.9B
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	0.18	55.4	≤3 s
2	0.23	57.7	≤3 s
4	0.32	51.3	≤3 s
8	0.76	46.2	≤3 s
16	1.14	37.8	≤3 s

Images per request	Response (s)	TPS (tok/s)	Target
1	0.18	55.4	≤3 s
2	0.32	53.9	≤3 s
4	0.65	51.1	≤3 s
8	1.31	46.9	≤3 s
16	2.78	40.8	≤3 s
32	6.13	32.6	>3 s
64	14.9	24.2	>3 s
128	41.4	15.6	>3 s
256	—	—	Failed

Cosmos3-Edge (3.9B parameters), served in BF16 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.18 s, then writes 55.4 tok/s.

With 128 images per camera at 720p, the answer starts after 41.37 s. With one 2K image, after 0.95 s.

At the default target of an answer starting within 3 s (720p, one image per camera), it keeps up with at least 16 cameras at once: it still met the target at the highest number measured, so the real limit is higher.