2025年年终全球人工智能状况报告_第1页
2025年年终全球人工智能状况报告_第2页
2025年年终全球人工智能状况报告_第3页
2025年年终全球人工智能状况报告_第4页
2025年年终全球人工智能状况报告_第5页
已阅读5页,还剩57页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

HighlightsEdition

ArtificialAnalysisStateofAI

ArtificialAnalysisStateofAI:2025Year-EndEdition

Thisreportincludesmodelreleasesuptotheendof2025.Forthelatestbenchmarkingresults,visittheliveArtificialAnalysiswebsiteatartificialanalysis.ai.

ArtificialAnalysisisaleadingandindependentAIbenchmarkingandinsightsprovider.WesupportengineersandcompaniestounderstandAIcapabilitiesandmakecriticaldecisionsabouttheirAIstrategy.

Ourdata,insightsandpublicationsaregroundedinourcomprehensivebenchmarkingofAItechnologiesandusecases.ThisincludeseverythingfromhourlyperformancetestingoflanguagemodelAPIstomillionsofvotesinourcrowd-sourcedarenas.

Ourpublicwebsite,artificialanalysis.ai,iswidelyreferencedbycompaniesleadinginnovationinAI.Todiscussthisreport,ourpublications,orourservices,pleasegetintouchatcontact@artificialanalysis.ai.

ArtificialAnalysisPremiumInsights:ComprehensiveAImarketintelligenceandinsightsforenterprisedecisionmakingfromtheleadingindependentbenchmarkingcompany

AICapabilityGuides

AIMarketIntelligence

EnterpriseAgentsGuide

Discoverhowagentsarereshapingproductivityanddeploymentacrossindustries

Additionalguideslaunchingsoon

Newguidesareaddedregularly,withafocusonhighprioritycapabilities

ModelDeploymentGuide

Comparemodels,inference

providers,andhardwarewithspecificbenchmarks

Q

QuarterlyStateofAIReports

StayaheadofAImarket

developmentswiththedefinitive

quarterlyupdate,incl.Chinareport

ThisReport

AIAdoptionSurvey

Gainreal-worldadoptioninsights

fromthosebuildinganddeployingAI

QuarterlyAIWebinars

ConnectthelatestAImarketintelligencetoyourstrategicpriorities

Trustedby

theleading

AIindustry

players,

mediaand

institutions

T⃞Techcrunch

WSJWentureBeat

CMetaGoogle

EntitiesthathavepubliclyreferencedArtificialAnalysis

AIStrategySupport

AIBenchmarkingSupport

AIDatabooks&APIAccess

Accesstheindustry’smost

comprehensiveAIperformanceandcostdata

AILaunchSupport

StrengthenyourAIlaunchwith

trustedperformancemetrics,brandassets,andindependentvalidation

AICustomBenchmarking

Evaluateandcomparemodels,

chips,andprovidersthroughour

customindependentbenchmarking

LeadersAIStrategyGuide

EquipyourleadershipteamtoharnessAIeffectivelyat

organizationalscale

AppliedAITrendsWorkshops

Engageyourteamsinaninteractive90-minutedeep-diveonthemost

importantAItrends

BespokeSupport

AccelerateyourAIstrategywith

expertsupportonplanning,

architecture,andimplementation

Jointheworld’sleadingAIlabsandenterprises

withsubscriptionsstartingfrom$3Kperquarter

subscriptions@artificialanalysis.ai

ThisHighlightsversionoftheQuarterlyStateofAIReportisalimitedversion.ThefullreportisavailabletosubscribersofourPremiumInsightsSubscription

IncludeseverythingintheHighlightsVersionplus:

Newlanguagemodelreleasecoverageandanalysis

(incl.analysisofleadingopenweightsoptions)

Modeltrendsanalysisoutliningemergingtrendsforlanguagemodelsacrosspricing,performanceandfeatures

Agentscoverageincludinganalysisofkeyagentcategories,use-casesandimplicationsforreal-worlddeployment

Imagegenerationmodelsandtrends(incl.texttoimageandimageediting)

Videogenerationmodelsandtrends(incl.texttovideoandimagetovideo)

Speechmodelsandtrends(incl.texttospeech,speechtotextandnativespeechtospeechmodels)

Emergingmarkettrendsforaccelerators,includingdetailedanalysiscomparingNVIDIAH100,H200andB200

Highlightsversion(thisversion)FullVersion(PremiumInsightsSubscription)

IndustryoverviewandmarketmapofkeyplayersandstrategiesacrosstheAIvaluechain

OverviewoffrontiermodelsrankedbytheArtificialAnalysisIntelligenceIndexandoverviewofemergingtrends

Synthesisofemergingtrendsforimage,videoandspeechmodelsandmarketmaps

SynthesisofemergingtrendsforacceleratorsincludingcasestudycomparingNVIDIAH100,H200andB200usingArtificialAnalysisSystemLoadTest

Feelfreetogetintouchwithusat

subscriptions@artificialanalysis.ai

tolearnmoreabouttheArtificialAnalysisPremiumInsightsSubscription

ArtificialAnalysisStateofAI:

2025Year-EndEdition

Justseveralmonthsago,inaletterjustlikethisone,weproclaimedthatrumorsofAIprogressslowinghadbeengreatlyexaggerated.Inearly2026,theideathatwewouldstartaletterlikethatseems

ridiculous.

Atthestartof2025,codingagentsdidn'texist.Bytheendoftheyear,theprofessionofsoftwareengineeringhadchangedforever-from

copy-pastingcodeintoChatGPTandCursorChattoinstructingagentsthatworkautonomouslyforseveralminutesatatime.Weexpect2026tobetheyearofagentsforeverythingelse.

Therewasnoconsolidationoftheracein2025–competitiononlyintensified,contributingtothecostofeverylevelofintelligence

continuingtofallconsistently.Progresswasdrivenbylabsscalingreinforcementlearning,focusingonlargesparsemixture-of-expertarchitectures,thearrivalofBlackwellhardwareandmore.

ProducedbyArtificialAnalysis,theleadingindependentAI

benchmarkingandinsightsprovider,this2025StateofAIReportis

designedtoinformproduct,engineeringandinvestmentdecisionsinanincreasinglyAI-nativeworld.

Formoredetails,contactusat

founders@artificialanalysis.ai

6

—MicahHill-SmithandGeorgeCameron,FoundersofArtificialAnalysis

Contents

1.IndustryOverview

OverviewofmarketmovementsandtrendsbykeyplayersintheAIindustry

2.LanguageModels

Trendsinfrontierlanguagemodels,including

increasingagenticintelligence,costandefficiencyimprovements

3.ImageandVideo

TrendsinfrontierimageandVideo

includinganoverviewoftheleadingmodelsinArtificialAnalysisImageandVideoArenas

4.SpeechandAudio

Trendsacrossnewspeechandmusic

modelsandanoverviewofnewandleadingmodelsintheArtificialAnalysisSpeechArena

5.Accelerators

OverviewoftheAIacceleratormarketincludingmarkettrends,available

acceleratorsandverticalintegrationbyselectchipmakers

01

8

01.IndustryOverview

Agentstakeoff

2025markedtheshiftfromsingle-query

Nativespeechtospeech

modelsgiverisetovoice

agents

TheAIindustrybecomesmore

contested

TheAIlandscapediversifiedsignificantly

in2025,withanexpandinglistof

companiesreleasingmodels.In2026,

weexpecttheracewillcontinueto

workloadstomulti-turnagentictasks.

2025sawamassiveimprovementin

speechtospeechqualitywiththe

developmentofnativeaudioreasoning

models,layingthefoundationsforvoice

agents

Codingagentsledearlyadoption;2026

willlikelyexpandagentsintobroader

enterpriseworkloads

becomemorecompetitive,notless

Reasoningmodelsbecomethe

statusquo

Imageeditingandvideogenerationgomainstream

5majortrendsshapedtheAIindustryin2025

Imageeditingandvideogeneration

reachedmainstreamviability,with

modelslikeGemini2.5Flash(Nano

Banana)deliveringstep-change

qualityimprovements

Atthestartof2025,OpenAI’so1

wastheonly‘reasoning’model,

however2025sawallAIlabsdevelop

reasoningmodelsthatnowoccupy

thespotsforthemostintelligent

models

01.IndustryOverview

GooglecontinuestobepositionedasAI'smostverticallyintegratedplayerfromTPUacceleratorstoGeminiapplications,spanningacrosstheentireAIvaluechain

NON-EXHAUSTIVE

KeyPlayersintheAIValueChain

ClassificationsareindicativeanddeterminedbasedonarangeoffactorsincludingmarketshareandstrengthofofferingNopresenceStrongpresence

Applications

Foundation

models(firstparty)

Cloud Inference(firstparty)

AcceleratorHardware

OpenAI

Anthropic

GGoogle

Microsoft

SambaNova

Amazon

DeepSeek

Snowflake

Databricks

Perplexity

Cerebras

Together.ai

Mistral

Alibaba

Cohere

NVIDIA

Meta

IBM

Groq

AMD

xAI

Fireworks

Nebius

9

Source:Companywebsite

01.IndustryOverview

TheAIlandscapeisbecomingincreasinglycompetitive,withnewinternationallabsenteringtheracein2025,thoughUSandChinafirmlylead

Keyplayerswithfirst-partymodelsbymodalityNON-EXHAUSTIVENomodelExistingmodel

SouthKorea

China

Other

Video

UnitedStates

OpenAI

Google

Meta

xAI

Anthropic

Microsoft

Amazon

NVIDIA

Adobe

ElevenLabs

Perplexity

Ai2

Midjourney

IBM

Alibaba

Tencent

Xiaomi

Baidu

DeepSeek

Bytedance

MiniMax

Z.ai

Kuaishou

MoonshotAI

LG

NAVER

Korea

ktTelecom

Upstage

Mistral

AI21labsAI21Labs

Cohere

MBZUAI

Language

Speech

Image

10

Source:Companywebsite

01.IndustryOverview

OpenAIbeganandended2025withthemostcapablelanguagemodel,buttheirleadisnarrowerthanever

FrontierLargeLanguageModel(LLM)IntelligencetillJan2026

ArtificialAnalysisIntelligenceIndexv4.0incorporates10evaluations:GDPval-AA,²-BenchTelecom,Terminal-BenchHard,SciCode,AA-LCR,AA-Omniscience,IFBench,Humanity'sLastExam,GPQADiamond,CritPt

GPT-5.2(xhigh)

Gemini3Pro

Preview(high)

Grok4

Claude4.5Opus(Reasoning)

o1

DeepSeekV3.2(Reasoning)

Llama4Maverick

G

TheintelligencefrontierisnowfiercelycontestedbetweenOpenAI,Anthropic,andGooglewithincreasingcompetitionfromlabsfromChina.MetahasrestructureditsAIeffortsandhasnotreleasedanewmodelsinceApril2025

11

Source:ArtificialAnalysisindependentbenchmarking

12

01.IndustryOverview

Efficiencyimprovedthroughmodelscaling…howeverlargerreasoningmodelsandmore

andhardware/softwareoptimizations…agenticworkloadsmeancomputedemand

continuedtoincrease

GPT-4levelintelligenceisnow100xcheaperthanoriginalGPT-4

Newapplicationscontinuetodemandmorecompute:asingledeepresearchquerycancost>10xanoriginalGPT-4query

A.SmallerModelsandSparsity

Algorithmicandtrainingdataimprovements

C.HardwareEfficiency

Nextgenerationacceleratorsoffer

B.SoftwareEfficiency

haveallowedsmaller

modelstogetsmarter

Inferenceoptimizations(e.g.FlashAttention)

improveefficiency

~1/3x

compute

~1/10x

compute

morecompute

efficiency

~1/3x

costs

~20x

requests/use

F.AIAgents

~10x

Agentschainmultiple

tokens/query

~5x

requeststoLLMsto

completetasks

E.ReasoningModels

compute/query

Significantincreasein

outputtokenswhen

models‘think’before

answering

autonomouslyacross

D.LargerModels

longconversations

todemandhigher

parametercountsfor

greaterintelligence

Scalinglawscontinue

Figuresarehighlyindicativeandservetoillustratethedirectional

impactofeachfactorimpactingcost

02

14

02.LanguageModels

2025wasdefinedbythereasoningparadigm,drivingsignificantintelligencegains,fallingcosts,andtheriseofagenticAI,asopenweightsandgloballabsnarrowedthegapwithUSfrontier

Keythemes

A.2025sawasignificantincreaseinmodelintelligence,drivenbya

paradigmshifttowardsreasoning

modelsthat‘think’beforeanswering

Byend-2025,OpenAI,Anthropic,andGoogleledtheintelligencefrontierwithreasoning-firstmodelsthat‘think’beforeanswering-markingaclearbreakfromearly2025,whennon-reasoningmodelsheldthe

topspotsasthemostintelligentmodels

Atthesametime,thereasoningparadigmmateriallyexpandedaverageworkloadsizeasmodelsgeneratedfarmoreoutputtokenswhen‘thinking’,whiledrivinghigherperformanceacross

general/scientificreasoning,long-horizonagentictasks,andcoding

B.2025markedtheriseofagenticAI,withmodelsincreasinglyexecuting

long-horizontasksend-to-end

Agentsevolvedfromtargetedusecases(e.g.deepresearch)togeneralizedtools,withfrontiermodelsnowreliablyorchestratingmulti-stepworkflowsacrossdomains

Toolcallingtrainingisnowuniversalwithmostmodelsreleasedin2025havingbeenpre-trainedandRL-optimizedforagentictaskexecution

Longhorizoncodingtaskswerethelargestbeneficiariesofimprovementsinagenticworkflowwithaclearproliferationofcodingagentsbeingreleasedin2025bysmallplayersandincumbents

C.2025witnessedademocratizationoffoundationmodels,thoughtheUSandChinamaintainasignificantlead

AIlabsfromacrosstheworld,includingEurope,MiddleEast,andAsiacontinuedtoreleasecompetitivefoundationmodels,howeverfrontiercapabilitiesremainconcentratedaroundUS(OpenAI,Anthropic,

Google)andChina(MoonshotAI,Z.ai,DeepSeek,Minimax)

WhiletheUSlabscontinuetoleadthedevelopmentofproprietaryfrontiermodels,Chineselabscontinuetoreleasefrontieropenweightsmodels

D.2025sawnewopenweightsmodelscontinuetokeeppacewithproprietarymodelsinintelligence,howeverthe

frontierisheldbyproprietarymodels

In2025,theopenweightsecosystemexpandedandbytheendof2025,themostcapableopenweightsmodelswereincreasinglyfromChineselabs

Throughout2025,openweightsmodelsbroadlykeptpacewithproprietarymodels,butproprietarymodelsstillledoverallintelligence

E.Thecostofintelligencefell

significantlyforo1-levelintelligence

Pricepertokenfell128xfortheo1-levelintelligencethat2025beganwith–drivenprimarilybysmallermodelsachievinghigherlevelofintelligence,aswellassoftwareandhardwareoptimizations

02.LanguageModels

A.Asattheendof2025,OpenAI,xAIandAnthropicleadfrontierintelligencewiththeirlatestreasoningmodelswithasignificantgaptothenextsetofAIlabs

LeadingLargeLanguageModels(LLMs)

ReasoningModel

ArtificialAnalysisIntelligenceIndexv4.0incorporates10evaluations:GDPval-AA,²-BenchTelecom,Terminal-BenchHard,SciCode,AA-LCR,AA-Omniscience,IFBench,Humanity'sLastExam,GPQADiamond,CritPt

•ThereasoningparadigmthatgaveOpenAIadecisiveleadatthestartof2025hasnowbeenadoptedbynearlyeverymajorlab,compressingOpenAI’sleadinoverallmodelintelligence

•DeepSeekR1releasedatthestartof2025,markedaturningpointasthefirstopenweightsreasoningmodelchallengingOpenAI’slead,trainedusingnoveltechniquesforpre-trainingandreinforcementlearning

•OpenAIstillleadswithGPT-5.2(xhigh),butcompetesinanincreasinglycrowdedfrontierwhereAnthropic,Google,xAI,andChineselabshaveallreleasedcompetitivereasoningmodels

15

Source:ArtificialAnalysisindependentbenchmarking

02.LanguageModels

A.Modelsreleasedin2025pushedtheintelligence-costfrontier:organizationscannowaccesshigherintelligenceatequivalentpricepoints,orequivalentintelligenceatmateriallylowercost

Intelligencevs.CosttoRunIntelligenceIndex(ParetoFrontiers)

ArtificialAnalysisIntelligenceIndexv4.0incorporates10evaluations:GDPval-AA,²-BenchTelecom,Terminal-BenchHard,SciCode,AA-LCR,AA-Omniscience,IFBench,Humanity'sLastExam,GPQADiamond,CritPt;CosttoRunIntelligenceIndex;Paretofrontierovertime

16

Source:ArtificialAnalysisindependentbenchmarking

17

02.LanguageModels

A.DeepDive:Largermodels(measuredbytotalparametercount)achievereliablyhigher

AA-Omniscienceaccuracyscores…

AA-OmniscienceAccuracyvs.TotalParameters(OpenWeightsModels)

AA-OmniscienceAccuracy;SizeinParameters(Billions)

18

02.LanguageModels

A.DeepDive:…buthallucinationrateislesscorrelatedwiththesizeofthemodel,indicatinggreaterimpactofothertrainingdecisions

AA-OmniscienceHallucinationRatevs.TotalParameters(OpenWeightsModels)

AA-OmniscienceHallucinationRate;SizeinParameters(Billions)

19

02.LanguageModels

B.2025wastheyearcodingagentsbegantowork;2026willbetheyear'agentsforeverything’begintowork

GPT-4released,enabling

seemingly-intelligentchat-

basedAIinterfacesandbroad

globaladoption

Reasoningbecamethe

standardinfrontiermodels,

andcodingagentsbeganto

work

20222023202420252026

ChatGPT(poweredbyGPT-

3.5)launched,bringing

LLMsintobroadpublicuse

GPT-4oexpandedmultimodal

functionalityenablingmodelsto

processandgeneratecontent

acrossdifferentmodalities(e.g.

image,video&speech)

Agentssteadilybegintotakeonawiderrangeofworktasks

?

02.LanguageModels

B.DeepDive:Asweshifttowardagenticworkflows,moreoutputtokensdoesn’ttranslatetohigherintelligence;intelligenceismoredrivenbyusingvarioustoolseffectively

GDPval-AA:ELOvs.TotalTokenUsage

GDPval-AAELO;TotaltokensusedtorunGDPval-AA

Amongfrontiermodels,Google'sandAnthropic’sleadingmodelsrepresenttheParetofrontierfortokenefficiencyinlonghorizonagentictasks

20

Source:ArtificialAnalysisindependentbenchmarking

02.LanguageModels

C.BeijingisemergingasahuboffrontierAIstartupactivity;Established“BigTech”

companiesaremoregeographicallydispersedwithnosinglenexusoftechinnovation

1

2

3

4

1Beijing

China’sleadingAIresearchcenter,combiningtopuniversities(Tsinghua,Peking),theBeijingAcademyofArtificialIntelligence,theworld'slargestconcentrationofAI

scientists,andZhongguancunScienceParktodominatefoundationalresearch

lilByteDance

⃞MoonshotAl

2Shanghai

AIhubfeaturingShanghaiFoundationModelInnovationCenter(China'sfirstandlargestfoundationmodelincubator),theShanghaiAILaboratory,andgovernmentinitiatives

targetinga$55billionAIindustrywithstrongsemiconductormanufacturingsupport

2Hangzhou

RisingAIhubpioneeringsmartcityapplicationsthroughAlibaba'sCityBrainplatform

andadedicated3.43sqkmAITownwithfull5Gcoverageforresearchanddevelopment

E2-amnceepseek

4Shenzhen

China'sAIhardwareandroboticsmanufacturingcapital,leveragingitsworld-class

electronicssupplychainandgiantslikeHuawei,Tencent,andDJIforAIdevelopment

Tencen广腾讯

Source:Crunchbase,PressSearch

02.LanguageModels

C.Korea’sgovernment-backedSovereignAIInitiativehascatalyzedthedomesticAI

ecosystem,producingmultiplenear-frontierAIlabs

NON-EXHAUSTIVE口

ShortlistedintheNationalInitiative

DescriptionofkeyKoreanmodellabs

⃞LGAIResearch

ipstage荨

Korea

Telecom

DescriptionAIresearcharmofLG

Koreantelecomgiant

AIstartupfoundedin

SouthKorea’smost

Establishedgame

Foundedin1981,KTis

Seoul-basedAI

Group.LGAIResearch

foundedin1984that

2020bySungKim(ex-

widelyusedsearch-

developerbehindhits

oneofthecountry’s

startupdeveloping

hasbeentheclear

leadsthelocal

NaverAIlead)that

engineandinternet

likeLineageandGuild

largesttelecom

proprietaryLLMand

front-runnerinthe

wirelessmarket.

hasquicklybecomea

conglomeratefounded

Wars,whoseAIarm

playersandprimarily

multimodalmodels.

KoreanAIraceforthe

Developinglanguage

powerhousein

in1999.Naverisnow

focusesonin-game

intendstointegrateits

Motifemphasizes

pastyear,consistently

modelsaspartof

Korea’ssovereignAI

applyingAIacross

andindustrialAIuse

AIstrategyinto

fullyhomegrown

leadinginintelligence

broaderdigital

racewithstrong

search,cloud,and

cases

existingproductsfor

architecture

benchmarks

transformationefforts

funding

consumerproducts

enterpriseclients

CompanyType

Largeconglomerate

Largeconglomerate

Startup

Largeconglomerate

Mediumsizedpubliccompany

Largeconglomerate

Startup

LatestK-EXAONE,a236B

A.XK1,a500Bopen

SolarOpen,a100B

HyperCLOVAXSEED

Varco-vision-2.0,a

Mi:dmK2.5Pro,a

Motif-2-12.7B,a

modelopenweightsreasoning

weightshyperscale

openweights

Think,a32Bopen

1.7Bimagegeneration

proprietaryreasoning

smallopenweights

model

reasoningmodel

reasoningmodel

weightsreasoningmodel

model

model

model

KoreanNationalSovereignAIInitiative

TheKoreanNationalSovereignAIInitiativeisagovernment-backed,nationwidecompetitionthatincentivizesdomesticmodeldevelopmentthroughamulti-stageelimination

process.Theinitiativeshortlistsnationalchampions,with

winnersreceivingdirectgovernmentfundingandguaranteedaccesstolarge-scaleGPUcapacity.

InAugust2025,5companieswereselectedinthefirststage…

InJanuary2026,3companieswereshortlistedinthesecondstage…

⃞LGAIResearch

iypstage

1moreshortlistedcompanywillbeannouncedinthecomingmonths…

⃞LGAIResearch

Source:ArtificialAnalysisindependentbenchmarking

02.LanguageModels

D.OpenAI’sfirstopenweightslanguagemodelsinceGPT-2pushedthefrontierforopenweightsmodels,buttheopen-to-proprietarygapremainssteady

LeadingLanguageModelsbyLicenseType,OverTime

ArtificialAnalysisIntelligenceIndexv4.0incorporates10evaluations:GDPval-AA,²-BenchTelecom,Terminal-BenchHard,SciCode,AA-LCR,AA-Omniscience,IFBench,Humanity'sLastExam,GPQADiamond,CritPt

GPT-5(high)

GPT-5.2

(xhigh)

OpenAIo1

OpenAIo1-preview

DeepSeekR1

OpenAIo3

DeepSeekR10528

\

gpt-oss-120B

(high)

KimiK2Thinking

KimiK2.5

23

Source:ArtificialAnalysisindependentbenchmarking

02.LanguageModels

E.Moreefficientmodelarchitecturecombinedwithsoftwareandhardwareefficiencies

helpeddrivedownmodelcosts-pertokenpricingfell128xforo1-levelintelligence

LanguageModelInferencePricebyIntelligenceCategory,OverTime

BlendedInput/OutputTokenPrice(USDperMTokens),ArtificialAnalysisIntelligenceIndexv4.0

GPT-4o1-previewo1

GPT-5.2(xhigh)

o3

GPT-3.5Turbo

Whilefrontierpricinghasdeclinedacrosssuccessiveintelligencecategories(fromGPT-4too1toGPT-5.2),thesereductionsaregradualandstepwise,incontrasttothecostdeclineatequivalentintelligencelevels

24

Source:ArtificialAnalysisindependentbenchmarking

03

03.ImageandVideo

MajorimprovementstobothImageandVideocamein2025,includingsupportformulti-modalinputs(imagetovideo,imageediting)andoutputs(videowithaudio)

KeyThemes

TexttoImageImprovesinQuality

Texttoimagemodelshaveimprovedsubstantiallyinquality,withGPTImage1.5(leaderatEOY2025)~150ELOpointshigherthanFLUX1.1[pro]Ultra(leaderatEOY2024)

ProgressinopenweightsimagemodelshasslowedasmajorlabssuchasOpenAIandGooglehaveenteredthespace.ThehighestrankingopenweightsmodelatEOYwasQwenImage2512,ranking#12intheTexttoImageLeaderboard

ImageEditingModelsLaunched

Instructionbasedimageeditingmodelsgainedpopularity,withthelaunchesofOpenAI’sGPT-4oImage,andGoogle’sNanoBanana(Gemini2.5Flash)drivingalargeincreaseinusageandmindshare

Multi-imageinputforimageeditingbecamecommon,withmodelssuchasNanoBananaProandQwenImageEditenablingmoreprecisecontrolofoutputimages

ImageGenerationmodelsbecameincreasinglygeneralized,supportingbothtexttoimageandimageeditinge.g.,theFLUX.2family,andSeedream4.5supportbothtexttoimage,andimageeditingmodalities

Videomodelsbreakintothemainstream

Videomodelssawabreakthroughinquality,withRunwayGen-4.5(leaderatEOY2025)~200ELOpointshigherthanOpenAI’sSora(leaderatEOY2024)

FocusonImagetoVideodrovestrongadoption,withusersabletocontrolvideogenerationswithmoregranularityandabletomaintaincharacterreferencesacrossshots

Openweightsvideomodelslaggedbehindproprietaryalternatives,withLTX-2ProrepresentingtheSOTAforopenweightsvideogeneration,ranking29thinTexttoVideoand28thinImagetoVideooverall

VideowithAudiostartswithVeo3

Veo3releasedinMay2025wasthefirsthighquality,mainstreammodelthatnativelysupportedaudiogenerationaspartofavideomodel,drivingstrongadoption

VideolabshavequicklyfollowedwiththeirownVideowithAudiomodels,suchasOpenAI’sSora2,Lightricks’LTX-2,Alibaba’sWan2.6,andByteDance’sSeedance1.5pro

ChinamaintainsparitywithUSinmedia

generationmodels

ChineseandUSlabscontinuetobeatparityforimagegenerationmodelswithByteDance’sSeedream4.5competitivewithGoogle’sNanoBananaPro,andOpenAI’sGPTImage1.5

ChineseandUSlabscontinuetobeatparityforvideogenerationmodelswithKling2.5TurbocompetitivewithVeo3.1andRunwayGen-4.5

03.ImageandVideo

Unlikeinlanguagemodels,smallermediagenerationfocusedAIlabshavecontinuedtocompetewithlargerlabswhohaveawiderbreadthofmodalitycoverage(1/2)

Keyplayersofferingimageand/orvideomodels(LabswithBroadFocus)

Existingmodel

Nomodel

Includespubliclyavailablemodelsineachmodalityreleasedinthelastyearbylabsthatdevelopbothlanguageandmediagenerationmodels

NON-EXHAUSTIVE

Modalities

OpenAI

GGoogle

ByteDance

MiniMax

Alibaba

Meta

Tencent

Amazon

Baidu

xAI

StepFun

NVIDIA

Continuesonnextpagefor

mediagenerationfocusedlabs

A.TexttoImage

B.ImageEditing

C.MultiImageEditing

D.TexttoVideo

E.ImagetoVideo

F.MultiImagetoVideo

G.VideowithAudioOutput

H.VideowithAudioInput

I.VideoEditing

27

03.ImageandVideo

Unlikeinlanguagemodels,smallermediagenerationfocusedAIlabshavecontinuedtocompetewithlargerlabswhohaveawiderbreadthofmodalitycoverage(2/2)

Keyplayersofferingimageand/orvideomodels(LabswithMediaGenerationFocus)

Existingmodel

Nomodel

Includespubliclyavailablemodelsineachmodalityreleasedinthelastyearbylabsthatdeveloponlymediagenerationmodels

NON-EXHAUSTIVE

Kuaishou

Modalities

Runway

Adobe

BlackForest

Lab

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论