← sources

DeepSeek has Two Numbers for Active Parameter Count. How?

source: reddit (reddit:LocalLLaMA)published 2026-11-11 16:47ingested 2026-11-12 02:42

retained text

Hi. It is frequently mentioned that DeepSeek v4.1 Flash has 8B "input" active parameters (aka prefill), and 16B output/decoding active parameters.

I have only read this about DeepSeek, and particularly v4.1 Flash (1017). Is it the only model that is like this? 

Is there an ELI5-or-so explanation of how this works, and why it is important? And if it is, can we expect other models to implement the same idea? Which of those numbers is "more important", if any?


Thanks

document fetched 2026-11-12 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xtmrii/deepseek_has_two_numbers_for_active_parameter