kinda crazy that deepseek v4 Pro, a 1.6T total 49B active parameters model, is cheaper than minimax M3, a 428B total, 23B active parameters one
DeepSeek V4 Pro (max) stands out on the Intelligence vs Cost per Task chart at $0.04 per task for an Intelligence Index of 44, running a single Intelligence Index task over 20x cheaper than GPT-5.5 (xhigh, $0.99) and over 40x cheaper than Claude Opus 4.8 (max, $1.78)
Users praise the DeepSeek V4 Pro architecture as remarkable for its efficiency gains from CSA and HCA over MSA in delivering high intelligence at low cost.
you can think that it's just outputing less token because it's smarter in a way, but it's the opposite!
kinda crazy that deepseek v4 Pro, a 1.6T total 49B active parameters model, is cheaper than minimax M3, a 428B total, 23B active parameters one
@eliebakouch I guess CSA and HCA are that much more efficient than MSA! There’s so much going on in the DS4 architecture it really is remarkable that it all came together. I wonder if mHC was a key step in making all that compressed attention work.
@eliebakouch I think part of the difference comes from the fact that Minimax takes a larger margin