Reaction
DeepSeek’s strongest model is claimed to have fewer parameters than V3
A commentator suggests MiniMax is also using smaller models and links the shift to compute scarcity and reinforcement learning.
TLDR
On October 19, a commentator said DeepSeek’s strongest model had fewer total and much fewer active parameters than V3 did in December 2024. They suggested a similar shift at MiniMax and argued that scarce compute was pushing resources toward reinforcement learning. In a follow-up, they cited “3x speed” and cost matching 4.1 except for caching, and guessed 13 billion active parameters.
Combined views
18.9K
3 Sources, first seen ago
