Cosmos-Reason2-8B on Jetson AGX Thor — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Cosmos-Reason2-8B
Parameters	8.8B
Device	Jetson AGX Thor
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	0.49	15.8	≤3 s
2	0.55	15.0	≤3 s
4	0.66	13.9	≤3 s
8	0.91	11.8	≤3 s
16	1.87	9.6	≤3 s
32	3.08	6.4	>3 s
64	5.53	3.9	>3 s
96	8.50	2.7	>3 s
112	10.7	2.4	>3 s
120	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.49	15.8	≤3 s
2	0.70	15.4	≤3 s
4	1.26	15.2	≤3 s
8	2.61	14.6	≤3 s
16	5.95	14.0	>3 s
32	14.4	12.5	>3 s
64	40.4	10.3	>3 s
128	127.4	7.7	>3 s
256	437.1	5.1	>3 s
512	—	—	Failed

Cosmos-Reason2-8B (8.8B parameters), served in BF16 format with vLLM on Jetson AGX Thor, starts answering a single camera that sends one 720p image after 0.49 s, then writes 15.8 tok/s.

With 256 images per camera at 720p, the answer starts after 437.06 s. With one 2K image, after 1.91 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.