Cosmos3-Nano on Jetson AGX Thor — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Cosmos3-Nano
Parameters	16B
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	0.48	15.7	≤3 s
2	0.54	15.0	≤3 s
4	0.57	14.3	≤3 s
8	1.04	13.1	≤3 s
16	1.97	10.8	≤3 s
32	3.26	8.0	>3 s
64	5.64	5.1	>3 s
96	8.46	3.7	>3 s
112	9.80	3.2	>3 s
120	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.48	15.7	≤3 s
2	0.68	15.6	≤3 s
4	1.23	15.1	≤3 s
8	2.60	15.0	≤3 s
16	5.98	14.4	>3 s
32	14.5	12.5	>3 s
64	40.2	10.5	>3 s
128	127.2	7.7	>3 s
256	436.4	5.2	>3 s
512	—	—	Failed

Cosmos3-Nano (16B parameters), served in BF16 format with vLLM on Jetson AGX Thor, starts answering a single camera that sends one 720p image after 0.48 s, then writes 15.7 tok/s.

With 256 images per camera at 720p, the answer starts after 436.36 s. With one 2K image, after 1.88 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.