Cosmos-Reason1-7B on DGX Spark — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Cosmos-Reason1-7B
Parameters	8.3B
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	0.51	13.5	≤3 s
2	0.74	15.7	≤3 s
4	1.18	15.0	≤3 s
8	1.95	13.3	≤3 s
16	3.49	11.2	>3 s

Images per request	Response (s)	TPS (tok/s)	Target
1	0.51	13.5	≤3 s
2	1.03	13.6	≤3 s
4	1.99	13.6	≤3 s
8	4.05	13.2	>3 s
16	8.57	12.8	>3 s
32	19.2	12.1	>3 s
64	46.8	10.6	>3 s
128	—	—	Failed

Cosmos-Reason1-7B (8.3B parameters), served in BF16 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.51 s, then writes 13.5 tok/s.

With 64 images per camera at 720p, the answer starts after 46.76 s. With one 2K image, after 2.03 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.