Cosmos-Reason2-2B on DGX Spark — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Cosmos-Reason2-2B
Parameters	2.4B
Device	DGX Spark
Quantization	BF16
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	64
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	0.16	49.7	≤3 s
2	0.22	55.1	≤3 s
4	0.34	51.3	≤3 s
8	0.52	43.5	≤3 s
16	0.88	33.2	≤3 s
32	1.58	22.4	≤3 s
64	2.47	12.5	≤3 s
128	4.97	6.6	>3 s
192	7.50	4.4	>3 s
224	9.16	3.8	>3 s
240	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.16	49.7	≤3 s
2	0.30	48.8	≤3 s
4	0.57	47.2	≤3 s
8	1.14	43.9	≤3 s
16	2.45	38.0	≤3 s
32	5.54	30.2	>3 s
64	13.8	21.3	>3 s
128	40.0	13.6	>3 s
256	129.3	7.8	>3 s
512	—	—	Failed

Cosmos-Reason2-2B (2.4B parameters), served in BF16 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.16 s, then writes 49.7 tok/s.

With 256 images per camera at 720p, the answer starts after 129.32 s. With one 2K image, after 0.75 s.

At the default target of an answer starting within 3 s (720p, one image per camera), it keeps up with 64 cameras at once.