DeepSeek-V4-Flash
DeepSeekMoE expert parameters use FP4 precision; most other parameters use FP8.
usedinference servingin DeepSeek-V4-FlashDeepSeek
implementation detail · filed under optimization
A mixed-precision scheme using FP4 for MoE expert parameters and FP8 for most other parameters.
The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See training precision for the methods filed beside it.
Also called FP4 + FP8 Mixed.
2 quotes from 2 sources, one card per model, each opening with its strongest.
MoE expert parameters use FP4 precision; most other parameters use FP8.
FP4 + FP8 Mixed: MoE expert parameters use FP4 precision; most other parameters use FP8.