Here’s a concise technical-focused summary of what DeepSeek V4 (formal release mid‑July 2026) actually improves and enables
Core architectural upgrades • Hybrid CSA/HCA attention: Uses Compressed Sparse Attention (CSA) plus Heavily Compressed Attention (HCA) to make 1M‑token context practically affordable, cutting FLOPs and KV cache to a fraction of V3.2 at the same length. • Engram memory: Separates “long‑term memory” from active GPU cache with O(1) lookup, so large contexts behave more like built‑in retrieval rather than brute‑force dense attention. • Manifold‑Constrained Hyper‑Connections (mHC): A new residual design that con
评论
?
参与讨论