A one-agent test of DeepSeek V4.1 Flash reached 848 tokens per second on Nvidia’s GB300 NVL72, giving a narrow measure of response speed rather than the rack’s overall output.
A roughly month-old reading on the same rack showed one agent generating 596 tokens per second, a gap of 252 tokens per second between the two results. These measurements describe output from an individual session and do not establish the rack’s combined production across multiple sessions. An unchanged rack configuration does not identify which of model design, serving software or hardware accounts for the difference.
The model accepts text and image inputs and uses a mixture-of-experts architecture. Its 552 billion parameters include 16 billion active during decoding, the stage when it generates responses; its context window is one million tokens.
A separate low-latency test on InferenceX put V4.1 Flash at 1,205 tokens per second with a compiler from Together AI. The Agent Arena Oct. 19 ranking placed the model first in its Open Source category.
