vLLM runs on NVIDIA Vera Rubin: 7.8x GB200 throughput on MiniMax M3, Chinese open models supported day one

Why it matters

【Key points】

On October 9 the vLLM team said its open-source inference engine now runs on NVIDIA's next-generation Vera Rubin NVL72 rack, with daily container builds and day-one support for Chinese open models from DeepSeek, Kimi, GLM and MiniMax. Inferact, NVIDIA and Red Hat worked with the vLLM community on it.

On SemiAnalysis's AgentX agentic benchmark with MiniMax M3, a Rubin GPU delivers up to 7.84x the throughput of GB200 at matched interactivity, and 5.18x at 150 tokens per second per user. In MLPerf Inference v6.1, Qwen3-VL-235B-A22B on vLLM plus Dynamo reached up to 3.7x the throughput of GB300 NVL72.

【Verifiable】

Sources are the vLLM blog post and the @vllm_project thread on X (about 36K views). The 7.84x compares against the older GB200, not GB300, covers one model, involved NVIDIA in testing, and has no independent reproduction. The blog cites about 5x rack-level FP4 inference compute while the X thread says about 3.5x per GPU. Rubin kernels for Kimi K3 and DeepSeek-V4.1-Flash are not yet integrated, so those models run but are untuned.

【Our read】

The signal is less about Rubin's speed than that Chinese open-weight models have become the default workload for benchmarking new AI hardware: MiniMax M3 headlines NVIDIA's first public Rubin inference results, and the vision test uses Qwen. Open weights earn that spot on hardware day one; closed models cannot.

Per-token costs will keep falling without waiting for model makers to cut prices, but the gain lands first with overseas clouds and hosts that can buy the newest NVIDIA racks. Early, single-model, vendor-involved numbers like 7.8x show direction, not procurement math.