Cosmos-Reason2-2B on Jetson AGX Thor — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Cosmos-Reason2-2B
Parameters	2.4B
Device	Jetson AGX Thor
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	0.24	62.6	≤3 s
2	0.27	58.2	≤3 s
4	0.33	52.2	≤3 s
8	0.61	45.8	≤3 s
16	1.08	34.9	≤3 s
32	1.52	22.4	≤3 s
64	2.37	13.1	≤3 s
128	4.49	7.0	>3 s
256	10.4	3.7	>3 s
288	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.24	62.6	≤3 s
2	0.37	61.0	≤3 s
4	0.62	57.9	≤3 s
8	1.25	53.0	≤3 s
16	2.76	46.4	≤3 s
32	6.47	35.3	>3 s
64	17.7	24.2	>3 s
128	52.6	15.1	>3 s
256	176.3	8.6	>3 s
512	—	—	Failed

Cosmos-Reason2-2B (2.4B parameters), served in BF16 format with vLLM on Jetson AGX Thor, starts answering a single camera that sends one 720p image after 0.24 s, then writes 62.6 tok/s.

With 256 images per camera at 720p, the answer starts after 176.28 s. With one 2K image, after 1.02 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.