vLLM, the open-source inference engine, said on 9 October 2026 that it now runs on Nvidia’s Vera Rubin NVL72, and that on SemiAnalysis AgentX it measured MiniMax M3 at up to 7.84x the throughput per GPU of GB200 NVL72 at matched interactivity, and 5.18x under a 150 tokens-per-second constraint.
MiniMax said on X on 10 October 2026 that its M3 model “is now running on NVIDIA Vera Rubin NVL72 with vLLM”. The chart behind the 7.84x is labelled an InferenceX preview whose results “may change as validation and publication continue.”
The vLLM blog post, dated 9 October 2026, says Inferact, NVIDIA, Red Hat and the vLLM community have been bringing vLLM up on Vera Rubin NVL72 since the platform was announced, and that vLLM runs on it today with daily container builds.
It says Rubin builds on Blackwell’s architecture family, so vLLM’s Blackwell kernels run on it, and that vLLM can already serve models from DeepSeek, Moonshot AI, Z.ai and MiniMax on Rubin.
The post lists Rubin-tuned attention, GEMM and MoE kernels arriving through FlashInfer 0.7.0, a MiniMax Sparse Attention prefill kernel tuned for Rubin, and nightly images built with CUDA 13.4 and PyTorch 2.15.
vLLM’s X post of 10 October 2026 summed it up: “vLLM now supports NVIDIA Vera Rubin. The early results show more than 7.8x the throughput of GB200 on MiniMax M3 on AgentX.”
The three performance multiples the vLLM blog states, each against its own baseline: AgentX with MiniMax M3, per chip against GB200 NVL72 (an InferenceX preview), and MLPerf v6.1 with Qwen3-VL against GB300 NVL72. They are different measures and do not add up.
Business Pill · BENCHMARKS
A one-minute explainer of AI benchmarks: what a benchmark score measures and what it leaves out. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, the 7.84x is a point on a curve: vLLM’s own chart puts it at the highest interactivity GB200 reached, and the same chart shows 5.18x at 150 tokens per second per user. Both are per-chip figures from a SemiAnalysis InferenceX preview on one model, MiniMax M3.
What Was Posted on X
vLLM now supports NVIDIA Vera Rubin. The early results show more than 7.8x the throughput of GB200 on MiniMax M3 on AgentX.
@inferact, @NVIDIAAI, @RedHat_AI, and the vLLM community have been bringing vLLM up on Rubin since it was announced. Here is where things stand.
The vLLM project’s post of 10 October 2026 opening its thread: vLLM now supports Nvidia Vera Rubin, with early results above 7.8x GB200 throughput on MiniMax M3 on AgentX. The piece reads the full figures from vLLM’s blog.
MiniMax M3 is now running on NVIDIA Vera Rubin NVL72 with vLLM!๐
Thanks to @vllm_project, @inferact, @NVIDIAAI, @RedHat_AI for bringing MiniMax M3 to Rubin.
MiniMax’s official account confirming the same day that MiniMax M3 is running on Vera Rubin NVL72 with vLLM.
Where the 7.84x Comes From
The post says that on AgentX, vLLM running MiniMax M3 on Vera Rubin NVL72 delivers up to 7.84x the throughput per NVIDIA GB200 at matched interactivity, and 5.18x higher throughput under a 150 TPS constraint.
The chart it publishes is titled “MiniMax M3 428B Agentic Token Throughput per Chip vs. P90 Interactivity”, credits SemiAnalysis InferenceX, and is marked updated 8 October 2026. It plots tokens per second per chip against tokens per second per user, and both systems run Dynamo vLLM.
On that chart, the 5.18x is marked at 150 tokens per second per user. The 7.84x is marked at the right end of the GB200 curve, at the GB200 point labelled TP8, which is the highest interactivity GB200 reached in the plot.
The chart labels the Vera Rubin points as TP2 and 1xTP2+4xTP2 and the GB200 points as TP4 and TP8, so the two systems appear at different parallelism settings.
What Rubin Changes, per vLLM
The post gives per-GPU comparisons with GB200 NVL72, citing NVIDIA spec pages and developer blogs: 5x the NVFP4 FLOPS, about 2.4x the HBM bandwidth with the move from HBM3e to HBM4, and 1.7x the bidirectional NVLink bandwidth.
It adds that Rubin raises exponential throughput, used in softmax, by 2x for FP32 and 4x for BF16 and FP16 against GB200.
On memory locality, the post says CUDA 13.4 locality domains let vLLM split MoE weights so each set of SMs reads its nearest memory, and that in preliminary MiniMax M3 MoE-layer tests this gave on average 1.2x speedups in small-token forward settings, with communication time not included.
Per-GPU hardware ratios, Vera Rubin NVL72 against GB200 NVL72, as the vLLM blog of 9 October 2026 states them, citing NVIDIA spec pages and developer blogs.
The MLPerf Result
The post also cites MLPerf Inference v6.1, where it says Vera Rubin NVL72, running Qwen3-VL-235B-A22B with vLLM as the inference engine and Dynamo as the router, delivered up to 3.7x higher throughput than GB300 NVL72 across offline, server and interactive scenarios.
NVIDIA’s own MLPerf blog of 16 September 2026 gives the same 3.7x against GB300 NVL72 on Qwen3-VL. It separately says that on SemiAnalysis AgentX, Vera Rubin NVL72 delivered 30x better performance than GB300 NVL72 in preview testing, a different comparison from vLLM’s.
The Structural Read
The vLLM post lists Rubin’s hardware (5x the NVFP4 FLOPS, about 2.4x the HBM bandwidth of GB200 NVL72), Rubin-tuned kernels through FlashInfer and locality-aware MoE among its highlights.
On the software side, the post says Rubin runs vLLM’s existing Blackwell kernels because it builds on the Blackwell architecture family, which is how it says vLLM could serve DeepSeek, Kimi, GLM and MiniMax models on day 0.
As we read it, the gap between 5.18x and 7.84x on the same chart is the reason a Rubin multiple needs its speed target attached: on one chart, the ratio is 5.18x at 150 tokens per second per user and 7.84x at GB200’s highest-interactivity point.
“vLLM now supports NVIDIA Vera Rubin. The early results show more than 7.8x the throughput of GB200 on MiniMax M3 on AgentX.”
Three Implications
MODEL BUILDERS MiniMax says M3 now runs on Vera Rubin NVL72 with vLLM, and the vLLM post lists DeepSeek, Kimi and GLM models as served on Rubin as well.
INFERENCE PROVIDERS The vLLM post says nightly Rubin images are available today on vLLM’s Docker Hub, built with CUDA 13.4 and PyTorch 2.15.
BENCHMARK READERS The AgentX chart is labelled an InferenceX preview whose results may change, and the vLLM post calls the numbers an early look with more gains expected.
The analysis puts it this way: “Training builds the brain. Inference is the brain working. And the AI brain never sleeps.”
As we read it, vLLM’s Rubin results are measured the way that framing suggests: tokens per second per chip at a given tokens-per-second-per-user target, the two numbers an inference operator trades against each other.
What Is Not Established
The vLLM post calls this an early look and says further gains are expected as optimisation continues; the chart itself is labelled a preview whose results may change. The 7.84x covers one model, MiniMax M3, on one benchmark.
The vLLM post and chart we read do not state the numeric precision used in the AgentX runs, give a cost-per-token or power figure, or report the result on Kimi, GLM or DeepSeek models. The post thanks NVIDIA for early access to Vera Rubin NVL72 hardware.
Business Engineer Framework
The State of The Inference Economy
A Business Engineer analysis of how inference became the AI economy’s revenue engine.
vLLM says it now runs on Vera Rubin NVL72 and that, on SemiAnalysis AgentX with MiniMax M3, Rubin delivered up to 7.84x the per-chip throughput of GB200 NVL72 at matched interactivity and 5.18x at 150 tokens per second per user, in results labelled a preview that may change.
A note on sourcing. We read the vLLM blog post dated 9 October 2026, including the labels on its AgentX chart, the vLLM and MiniMax posts on X of 10 October 2026, and NVIDIA’s MLPerf blog of 16 September 2026, on 11 October 2026. Every performance figure is the vLLM project’s or NVIDIA’s own, and the AgentX results are labelled a preview. Nothing here is a forecast, and nothing here is financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
