Xiuyu Li
@sheriyuo
Researcher @StepFun_ai | Working on long-horizon tasks | Prev @RUC1937 | Opinions are my own
๊ฐ์
February 2026
1.7K ํ๋ก์ ์ค 12.4K ํฌ
So it really just means that V4 did not perform well during training from the start ๐ง
There are two hypotheses for the DeepSeek-V4's strange performance (as in, V4-Flash is about as good as we expected, but V4-Pro is disappointing given its scale):
1) failed pretrain
2) big difference in the RL/MOPD stage
Flash probably got multiple such iterations
๋ ๋ณด๊ธฐ
