vLLM Publishes DeepSeek-V4.1-Flash Performance on Vera Rubin NVL72
Posted As of
The vLLM blog published performance results for serving DeepSeek-V4.1-Flash on a Vera Rubin NVL72 rack. The 58x figure over GB300 NVL72 compares at a specific interactivity point and should not be read as an overall speedup.
What happened
- Per the official vLLM blog, the post was published on 2026-11-12.
- The benchmark framework is SemiAnalysis AgentX.
- (Official) At 387 P90 tokens/s per user, it reported 1.7M tokens/s per GPU, which it says is 58x GB300 NVL72.
- (Official) At peak, it presented 2.52M tokens/s per GPU and 1,810 tokens/s for a single user.
- (Third-party analysis) It noted the gap at peak throughput is 2.7x.
- The post is the third in a GB300 series and explains that it moves the same stack to Vera Rubin NVL72.
Our take
These are datacenter-class rack results that are hard to compare directly with setups individuals run on a Mac or PC. What matters for local users is mainly how inference engine optimization changes alongside hardware generations. Independent verification has not been confirmed, and model precision, batch configuration, and cost are not confirmed in the excerpt.
Sources and related posts
- vLLM 공식 블로그 Primary source