Cosmos-Reason2-8B on DGX Spark — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Cosmos-Reason2-8B
Parameters	8.8B
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	0.39	14.5	≤3 s
2	0.53	14.3	≤3 s
4	0.65	13.2	≤3 s
8	0.99	11.2	≤3 s
16	1.86	8.4	≤3 s
32	3.26	5.9	>3 s
64	6.06	3.4	>3 s
96	9.76	2.4	>3 s
104	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.39	14.5	≤3 s
2	0.72	14.2	≤3 s
4	1.39	14.1	≤3 s
8	2.77	13.7	≤3 s
16	5.90	13.0	>3 s
32	13.5	11.7	>3 s
64	33.2	9.6	>3 s
128	96.9	7.2	>3 s
256	314.6	4.8	>3 s
512	—	—	Failed

Cosmos-Reason2-8B (8.8B parameters), served in BF16 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.39 s, then writes 14.5 tok/s.

With 256 images per camera at 720p, the answer starts after 314.56 s. With one 2K image, after 1.71 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.