Reading List

DeepSeek researchers detail mHC, a new architecture they used to train 3B, 9B, and 27B models, finding it scaled without adding significant computational burden (Vincent Chow/South China Morning Post) from Techmeme RSS feed.