Qwen3-VL

NVIDIA's 3.7x Vera Rubin figure comes from Qwen3-VL on vLLM and Dynamo

MLPerf Inference v6.1 preview submissions put Vera Rubin NVL72 at up to 3.7x GB300 on Qwen3-VL and up to 2.5x on DeepSeek-R1, on two different inference frameworks. The four-rack 99% scaling result is an offline number.