谷歌：大模型Gemini技术报告 .pdf

1、本文档共65页，可阅读全部内容。
2、原创力文档（book118）网站文档一经付费（服务费），不意味着购买了该文档的版权，仅供个人/单位学习、研究之用，不得用于商业用途，未经授权，严禁复制、发行、汇编、翻译或者网络传播等，侵权必究。
3、本站所有内容均由合作方或网友上传，本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺！文档内容仅供研究参考，付费前请自行鉴别。如您付费，意味着您自己接受本站规则且自行承担风险，本站不退款、不进行额外附加服务；查看《如何避免下载的几个坑》。如果您已付费下载过本站文档，您可以点击这里二次下载。
4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等，请点击“版权申诉”（推荐），也可以打举报电话：400-050-0827(电话支持时间：9:00-18:30)。

Gemini:AFamilyofHighlyCapable

MultimodalModels

GeminiTeam,Google1

Thisreportintroducesanewfamilyofmultimodalmodels,Gemini,thatexhibitremarkablecapabilities

acrossimage,audio,video,andtextunderstanding.TheGeminifamilyconsistsofUltra,Pro,andNano

sizes,suitableforapplicationsrangingfromcomplexreasoningtaskstoon-devicememory-constrained

use-cases.Evaluationonabroadrangeofbenchmarksshowsthatourmost-capableGeminiUltramodel

advancesthestateoftheartin30of32ofthesebenchmarks—notablybeingthefirstmodeltoachieve

human-expertperformanceonthewell-studiedexambenchmarkMMLU,andimprovingthestateofthe

artineveryoneofthe20multimodalbenchmarksweexamined.Webelievethatthenewcapabilitiesof

Geminimodelsincross-modalreasoningandlanguageunderstandingwillenableawidevarietyofuse

casesandwediscussourapproachtowarddeployingthemresponsiblytousers.

1.Introduction

WepresentGemini,afamilyofhighlycapablemultimodalmodelsdevelopedatGoogle.Wetrained

Geminijointlyacrossimage,audio,video,andtextdataforthepurposeofbuildingamodelwithboth

stronggeneralistcapabilitiesacrossmodalitiesalongsidecutting-edgeunderstandingandreasoning

performanceineachrespectivedomain.

Gemini1.0,ourfirstversion,comesinthreesizes:Ultraforhighly-complextasks,Proforenhanced

performanceanddeployabilityatscale,andNanoforon-deviceapplications.Eachsizeisspecifically

tailoredtoaddressdifferentcomputationallimitationsandapplicationrequirements.Weevaluate

theperformanceofGeminimodelsonacomprehensivesuiteofinternalandexternalbenchmarks

coveringawiderangeoflanguage,coding,reasoning,andmultimodaltasks.

Geminiadvancesstate-of-the-artinlarge-scalelanguagemodeling(Aniletal.,2023;Brownetal.,

2020;Chowdheryetal.,2023;Hoffmannetal.,2022;OpenAI,2023a;Radford

该用户很懒，什么也没介绍

咨询Ta 进入空间

更多 >