tracked author

Zhenda Xie (解振达)

DeepSeek researcher and recurring DeepSeekMoE / Engram / V-series author. GitHub identifies the same handle as Researcher @deepseek-ai, Scholar verifies a deepseek.com email, DBLP links his earlier self-supervised learning work, and X bio matches Researcher @ DeepSeek AI plus foundation-model pretraining/scaling.

3 archived notes X: high GitHubX

Related Notes

按论文归档时间排序,展示该作者在本站已经出现的材料。

归档

DeepSeek V4.1 Flash: Pushing the Limits of KV Cache Compression

DeepSeek V4.1 Flash 将因果编码器与解码器分工、跨层稀疏 KV 复用、四位全局缓存和近似窗口重放组合起来,把长输入的预填充主干计算近似减半、全局 KV 降至每 token 890 字节,并在发布方评测中显著提升多项 agent 能力;端到端服务收益和近似重放的极端条件可靠性仍缺少充分公开验证。

待审阅 2026-09-10-deepseek-v4-1-flash-kv-cache-compression KV CacheSparse AttentionMultimodal Model