DeepSeek-V3 – 知乎

January 23, 2025

0. 引言本篇讲讲deepseek在MoE（Mixture-of-Experts）上的演进过程。DeepSeek是MoE稀疏模型的忠实玩家。主版本模型从DeepSeekMoE（V1）到 DeepSeek V3，一直坚持走MoE的技术路线，并且持续做出一些创新。本文参考paper并结合源码阅读，理解MoE的演进过程和具体实现。

https://www.zhihu.com/topic/607345252

DeepSeek-V3 – 知乎 <body> <h1> Array ( [0] => seepseek-v3 – 知乎 [1] => xeepseek-v3 – 知乎 [2] => ceepseek-v3 – 知乎 [3] => feepseek-v3 – 知乎 [4] => reepseek-v3 – 知乎 [5] => eeepseek-v3 – 知乎 [6] => dwepseek-v3 – 知乎 [7] => dsepseek-v3 – 知乎 [8] => ddepseek-v3 – 知乎 [9] => drepseek-v3 – 知乎 [10] => d4epseek-v3 – 知乎 [11] => d3epseek-v3 – 知乎 [12] => dewpseek-v3 – 知乎 [13] => despseek-v3 – 知乎 [14] => dedpseek-v3 – 知乎 [15] => derpseek-v3 – 知乎 [16] => de4pseek-v3 – 知乎 [17] => de3pseek-v3 – 知乎 [18] => deeoseek-v3 – 知乎 [19] => deelseek-v3 – 知乎 [20] => dee-seek-v3 – 知乎 [21] => dee0seek-v3 – 知乎 [22] => deepaeek-v3 – 知乎 [23] => deepzeek-v3 – 知乎 [24] => deepxeek-v3 – 知乎 [25] => deepdeek-v3 – 知乎 [26] => deepeeek-v3 – 知乎 [27] => deepweek-v3 – 知乎 [28] => deepswek-v3 – 知乎 [29] => deepssek-v3 – 知乎 [30] => deepsdek-v3 – 知乎 [31] => deepsrek-v3 – 知乎 [32] => deeps4ek-v3 – 知乎 [33] => deeps3ek-v3 – 知乎 [34] => deepsewk-v3 – 知乎 [35] => deepsesk-v3 – 知乎 [36] => deepsedk-v3 – 知乎 [37] => deepserk-v3 – 知乎 [38] => deepse4k-v3 – 知乎 [39] => deepse3k-v3 – 知乎 [40] => deepseej-v3 – 知乎 [41] => deepseem-v3 – 知乎 [42] => deepseel-v3 – 知乎 [43] => deepseeo-v3 – 知乎 [44] => deepseei-v3 – 知乎 [45] => deepseek0v3 – 知乎 [46] => deepseekpv3 – 知乎 [47] => deepseek-c3 – 知乎 [48] => deepseek-b3 – 知乎 [49] => deepseek-g3 – 知乎 [50] => deepseek-f3 – 知乎 [51] => deepseek-v2 – 知乎 [52] => deepseek-vw – 知乎 [53] => deepseek-ve – 知乎 [54] => deepseek-v4 – 知乎 [55] => deepseek-v3 ᰫ 知乎 [56] => deepseek-v3 &#u211; 知乎 [57] => deepseek-v3 &#i211; 知乎 [58] => deepseek-v3 ⏻ 知乎 [59] => deepseek-v3 ᾯ 知乎 [60] => deepseek-v3 &#8q11; 知乎 [61] => deepseek-v3 &#8w11; 知乎 [62] => deepseek-v3 ⁷ 知乎 [63] => deepseek-v3 ” 知乎 [64] => deepseek-v3 &#82q1; 知乎 [65] => deepseek-v3 — 知乎 [66] => deepseek-v3 &#821q; 知乎 ) </h1> </body>