MiMo-V2.6-Distill-Qwen-9B on DGX Spark — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	MiMo-V2.6-Distill-Qwen-9B
Parameters	9.4B
Device	DGX Spark
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	0.32	12.8	≤3 s
2	0.46	13.7	≤3 s
4	0.67	13.2	≤3 s
8	1.23	12.3	≤3 s
16	2.29	10.5	≤3 s
32	3.68	7.8	>3 s
64	6.23	5.1	>3 s
96	8.87	3.8	>3 s
104	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.32	12.8	≤3 s
2	0.62	12.8	≤3 s
4	1.22	12.8	≤3 s
8	2.44	12.7	≤3 s
16	4.97	12.6	>3 s
32	10.2	12.3	>3 s
64	21.9	11.7	>3 s
128	49.7	10.8	>3 s
256	—	—	Failed

MiMo-V2.6-Distill-Qwen-9B (9.4B parameters), served in BF16 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.32 s, then writes 12.8 tok/s.

With 128 images per camera at 720p, the answer starts after 49.72 s. With one 2K image, after 1.54 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.