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

Open this configuration in the VLM Inference Benchmark Explorer

Model	Cosmos3-Edge
Parameters	3.9B
Device	Jetson AGX Thor
Quantization	BF16
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	32
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	0.24	63.8	≤3 s
2	0.30	58.2	≤3 s
4	0.44	53.5	≤3 s
8	0.80	46.6	≤3 s
16	1.51	34.8	≤3 s
32	2.30	22.3	≤3 s
64	3.34	12.3	>3 s
128	5.88	6.3	>3 s
192	8.99	4.4	>3 s
208	9.85	4.1	>3 s
224	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.24	63.8	≤3 s
2	0.38	61.8	≤3 s
4	0.75	58.4	≤3 s
8	1.46	54.2	≤3 s
16	3.23	47.4	>3 s
32	7.46	36.4	>3 s
64	19.4	24.2	>3 s
128	56.1	15.3	>3 s
256	—	—	Failed

Cosmos3-Edge (3.9B 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 63.8 tok/s.

With 128 images per camera at 720p, the answer starts after 56.09 s. With one 2K image, after 1.34 s.

At the default target of an answer starting within 3 s (720p, one image per camera), it keeps up with 32 cameras at once.