---
format: "aidr-story-markdown/v1"
id: "ec5eddb48a5c61b6298fa285f598a20a45d19bc81d7ab389742a8be2dd7f0472"
canonical_url: "https://aidr.today/ec5eddb4?lang=en"
title: "vLLM Achieves 93.3x the Token Throughput of GB300 for DeepSeek Inference on Nvidia Rubin"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-16T23:38:12.000Z"
category: "Frameworks"
topics: ["deepseek","inference","framework"]
source_urls: []
summary: "The vLLM project reported that the DeepSeek-V4.1-Flash model running on Nvidia Vera Rubin NVL72 hardware achieves a token throughput 63.4 to 93.3 times higher than the GB300 NVL72, depending on the user rate. To optimize performance, vLLM implemented a Mooncake KV pool and sharded Engram tables over NVLink."
---

# vLLM Achieves 93\.3x the Token Throughput of GB300 for DeepSeek Inference on Nvidia Rubin

> [Open the canonical story](<https://aidr.today/ec5eddb4?lang=en>)

**Published:** 2026-11-16T23:38:12.000Z
**Category:** Frameworks
**Topics:** deepseek, inference, framework

## Summary

The vLLM project reported that the DeepSeek\-V4\.1\-Flash model running on Nvidia Vera Rubin NVL72 hardware achieves a token throughput 63\.4 to 93\.3 times higher than the GB300 NVL72, depending on the user rate\. To optimize performance, vLLM implemented a Mooncake KV pool and sharded Engram tables over NVLink\.

## Sources

_No safe source URL was recorded._

