DeepSeek被低估：Kimi和智谱旗舰模型基于其旧架构

There's been a meme spreading that DeepSeek is Not That Good at ML, "just god-tier infra". That's funny when both Kimi and Zhipu's flagships run on DeepSeek's older designs. Not because they're lazy. Because people like Jianlin Su looked and said "…would be hard to beat this".