YOCO back in spotlight: blog breaks down cross-layer KV sharing in DeepSeek-V4.1-Flash and Gemma 4
donglixp · x · 2026-09-21
- With the reported release of DeepSeek-V4.1-Flash (unconfirmed), its adoption of YOCO (You Only Cache Once) has put the architecture back in the spotlight.
- The author published a short technical note (PDF, in Chinese and English) covering YOCO and cross-layer KV sharing, including YOCO-U, YOCO-Sparse, and related designs in DeepSeek-V4.1-Flash, Gemma 4 E2B/E4B, and WeLM.
- The author notes possible omissions; the English version was AI-translated from the Chinese original.
More from Models
- TypeSafe's JEV model fully open with $5 free credit, powers 1-second 3D scene generation — tinyfool · 2026-09-21
- Prof. Tom Yeh Posts Kimi 3 Seminar Recording with Nathan Lambert on KDA Architecture — ProfTomYeh · 2026-09-21
- rxwei on Jev: Co-Designing SDKs and Models Unlocks a New Paradigm — rxwei · 2026-09-21
- Open models are ~10% of AI revenue, but the analysis used badly outdated data — xeophon · 2026-09-21
- Open-Source JevGraph Cuts KG Extraction Cost 3x vs DeepSeek With 2.9x Speedup — richie9830 · 2026-09-21
- In 2026, Thinking Machines' Inkling is the only truly 'open' US model, not OpenAI — Conscious-Neat-5015 · 2026-09-21