2026KimiK3技术报告英文版.pptxVIP

  • 0
  • 0
  • 约19.31万字
  • 约 10页
  • 2026-08-03 发布于北京
  • 举报

TECHNICALREPORTOFKIMIK3

ABSTRACT

WeintroduceKimiK3,a2.8TparameterMixture-of-Expertsmodelwith104billionactivatedparameters,nativevisioncapabilities,anda1-million-tokencontextwindow.KimiK3isbuiltonKimiDeltaAttention[64]andAttentionResiduals[58],whichimproveinformationflowacrosssequencelengthandmodeldepth.TogetherwithStableLatentMoE,whicheffectivelyactivates16of896routedexpertspertoken,andrefinedtraininganddatarecipes,theseadvancesyieldanapproximately2.5×improvementinoverallscalingefficiencyoverKimiK2[59].Pos

文档评论(0)

1亿VIP精品文档

相关文档