Qwen3.8-27B on Jetson AGX Thor — FP8, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Qwen3.8-27B
Parameters	27B
Device	Jetson AGX Thor
Quantization	FP8
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	1
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	2.90	8.6	≤3 s
2	4.87	8.2	>3 s
4	8.20	7.4	>3 s
8	10.6	5.8	>3 s
16	7.00	5.7	>3 s
32	10.9	3.7	>3 s
40	12.2	3.0	>3 s
44	12.8	2.9	>3 s
48	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	2.90	8.6	≤3 s
2	1.43	8.6	≤3 s
4	2.74	8.6	≤3 s
8	5.40	8.6	>3 s
16	11.3	8.4	>3 s
32	24.2	8.2	>3 s
64	55.5	7.8	>3 s
128	140.3	7.0	>3 s
256	403.6	5.8	>3 s
512	—	—	Failed

Qwen3.8-27B (27B parameters), served in FP8 format with vLLM on Jetson AGX Thor, starts answering a single camera that sends one 720p image after 2.9 s, then writes 8.6 tok/s.

With 256 images per camera at 720p, the answer starts after 403.62 s. With one 2K image, after 7.01 s.

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