# DeepSeek V4.1 Flash Logs 848 Tokens per Second per Agent on GB300 NVL72 By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/51765 Published: 2026-11-08T19:33:37.000Z Updated: 2026-11-08T19:33:37.000Z Section: Technology > A reading about a month earlier was 596 tokens per second per agent on the same rack. The figure measures one session’s output pace, not total system throughput. DeepSeek V4.1 Flash produced 848 tokens a second for one agent on Nvidia’s GB300 NVL72, up 252 over roughly a month on the same setup, a change relevant to AI capacity planning. The earlier reading was 596 tokens per second per agent. The rack setup remained the same, while the broader performance work involves model design, serving software and hardware; the comparison does not isolate the contribution of each. Per-agent decoding speed measures how quickly one session generates output. Aggregate throughput totals output across a system, so the two measures describe different scales of activity. V4.1 Flash is a multimodal mixture-of-experts model with 552 billion parameters in total. It activates 16 billion parameters during decoding, when it generates a response, and accepts text and image inputs. Its context window extends to 1 million tokens. Nvidia’s GB300 NVL72 places 72 GPUs in a single rack. In a separate [low-latency InferenceX benchmark](), Together AI’s compiler reached 1,205 tokens per second while running DeepSeek V4.1 Flash. ## Links in this article - [low-latency InferenceX benchmark](https://www.tokenpost.com/news/technology/49149)