这篇笔记详细的介绍了DeepSeek-V4.1-Flash 背后的跨层 KV 共享：YOCO，以及相关的变种架构
This is a AI post classified by Jev as AI research (a free resource), kept by the AI Radar because it carries real work, not commentary.
这篇笔记详细的介绍了DeepSeek-V4.1-Flash 背后的跨层 KV 共享：YOCO，以及相关的变种架构 1、YOCO 本身的架构非常简单易懂：把标准 decoder-only Transformer 拆成上下两部分。 - 前半部分是 Self-Decoder。 - 后半部分是 Cross-Decoder。 Prefill 阶段则完全不需要跑后面的 Cross-Decoder（这在现在的 agent 场景的超长输入中，确实是太实用了）。 2、YOCO-U：对前面的 Self-Decoder loop 三次后产生一份 Shared KV；在 Cross-Decoder 上的改动则是采用了 NOPE。 3、YOCO-Sparse：共享 Sparse Attention 的 Routing Index，省中省！ 4、DeepSeek-V4.1-Flash：Attention 大致为 YOCO/CED + CSA2 + SWA。 - YOCO/CED：减少 Prefill 计算 - CSA2：跨层复用 Global KV 与 Top-K - SWA：短期缓存与局部重建 5、Gemma 4 E2B / E4B 也用了类似 YOCO 的技术。后半段很多层不再计算自己的 K/V，只保留独立的 Q，并复用前面最后一个同类型 Attention 层的 KV。 6、WeL
Posted by 青稞社区 (3.2k followers) 2 h ago · 4 likes · 686 views · view the original post on X. Kept by the AI Radar as AI research. Tools mentioned: purshow_notes.
More AI work like this
- Impressive paper on building recursive self-improving agent harnesses. — @dair_ai
- JUST IN: Q* has been solved. — @yifanzhang_
- 这篇笔记详细的介绍了DeepSeek-V4.1-Flash 背后的跨层 KV 共享：YOCO，以及相关的变种架构 — @qingke_ai
- OpenAI researcher "Alisa Liu" had 57 interviews before joining OpenAI, and then She… — @0x0SojalSec
- UPDATE! — @BrianRoemmele
- HerHealthEval tests clinical LLMs across English, French, Arabic in six registers. — @biogerontology
- "JEPA-Anything: Learning Predictive Models across Different Worlds" — @askalphaxiv
- A while ago, I realized that I needed to actually learn AI fundamentals, not just… — @0xdoug
Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 35.3k posts from 5k X accounts over the last 14 days, 1.5k tools, 19 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-21 02:10 UTC. Full method.