Skip to content
QbitAI · WeChat

DeepSeek V4’s biggest omission

DeepSeek V4最大的遗憾

DeepSeek V4’s technical report omits Engram while listing mHC, CSA, HCA, Muon, and FP4. Engram was open-sourced by DeepSeek and Peking University in January, inserting lookup modules between Transformer layers 2 and 15; its 27B test raised MMLU by 3.4 and Multi-Query NIAH to 97.0%. The engineering signal is CXL pooling: 8 servers shared a 4TB memory pool with under 5% throughput loss.

Why it matters: The DeepSeek V4 technical report leaves out the Engram lookup module that Peking University open-sourced in January, which lifted MMLU by 3.4 points at 27B scale and hit 97.0% on multi-query needle-in-a-haystack.

Read the original ↗Export Markdown