NVIDIA's export of DeepSeek V4.1 Flash changes the numerical format of its routed expert components for deployment on Blackwell hardware. The source already used four-bit expert weights, so this conversion does not make the checkpoint smaller: NVIDIA reports an increase from approximately 476 GiB to 492 GiB because of the new scaling layout. Other components retain their source precision. The model card documents testing with vLLM and SGLang on GB300 hardware. Its benchmark results came from vLLM, while SGLang was checked for loading and generation, so the card gives operators a specific configuration to evaluate rather than a measured advantage over the source checkpoint.
Terms in this piece · Glossary
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
