Qwen3-VL-30B-A3B-Instruct on DGX B300 (1 GPU) — FP8, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Qwen3-VL-30B-A3B-Instruct
Parameters	30B
Device	DGX B300 (1 GPU)
Quantization	FP8
Image size	720p
Images per camera	1
Max Cameras (TTFT ≤ 3000 ms)	128
Engine	vLLM

Cameras	TTFT (ms)	TPS (tok/s)	Target
1	83	298.0	≤3000 ms
2	125	247.5	≤3000 ms
4	200	221.0	≤3000 ms
8	289	192.8	≤3000 ms
16	562	139.6	≤3000 ms
32	594	74.3	≤3000 ms
64	2067	37.4	≤3000 ms
128	2772	44.7	≤3000 ms
256	5397	35.6	>3000 ms
384	8019	28.2	>3000 ms
416	—	—	Failed

Images per request	TTFT (ms)	TPS (tok/s)	Target
1	83	298.0	≤3000 ms
2	177	295.4	≤3000 ms
4	200	289.9	≤3000 ms
8	329	279.4	≤3000 ms
16	694	262.7	≤3000 ms
32	1635	245.6	≤3000 ms
64	3509	221.2	>3000 ms
128	9232	188.5	>3000 ms
256	26497	142.7	>3000 ms
512	—	—	Failed

Qwen3-VL-30B-A3B-Instruct (30B parameters), served in FP8 format with vLLM on DGX B300 (1 GPU), starts answering a single camera that sends one 720p image after 83 ms (time to first token, TTFT), then writes 298 tok/s.

With 256 images per camera at 720p, the answer starts after 26,497 ms. With one 2K image, after 245 ms.

At the default target of a mean TTFT within 3,000 ms (720p, one image per camera), it keeps up with 128 cameras at once.