Nemotron-Nano-12B-v2-VL on Jetson AGX Thor — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Nemotron-Nano-12B-v2-VL
Parameters	13B
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	0.58	10.2	≤3 s
2	0.68	9.8	≤3 s
4	0.81	9.6	≤3 s
8	1.31	8.5	≤3 s
12	2.08	7.4	≤3 s
16	2.49	6.8	≤3 s
32	4.17	5.2	>3 s
64	7.35	3.6	>3 s
80	9.10	3.0	>3 s
88	10.1	2.8	>3 s
96	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.58	10.2	≤3 s
2	1.00	10.1	≤3 s
4	1.63	10.3	≤3 s
8	3.15	11.5	>3 s
16	6.03	9.9	>3 s
32	12.1	10.7	>3 s
64	26.2	10.2	>3 s
128	59.1	12.0	>3 s
256	—	—	Failed

Nemotron-Nano-12B-v2-VL (13B parameters), served in BF16 format with vLLM on Jetson AGX Thor, starts answering a single camera that sends one 720p image after 0.58 s, then writes 10.2 tok/s.

With 128 images per camera at 720p, the answer starts after 59.14 s. With one 2K image, after 1.36 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.