Artificial

AnalysisQ2

2025HighlightsReportFullreportavailabletoPremium

Access

subscribersStateof

AI:ChinaArtificialAnalysis

isaleading,andindependentAI

benchmarkingandinsightsprovider.Wesupportengineersandcompaniesto

understandAIcapabilitiesandmakecriticaldecisionsabout

their

AIstrategy.Ourdata,insightsandpublicationsaregrounded

in

our

comprehensivebenchmarkingofAItechnologiesandusecases.Thisincludeseverythingfrom

hourlyperformancetestingoflanguagemodelAPIsto

millionsofvotes

inourcrowd-sourcedarenas.Our

publicwebsite,artificialanalysis.ai,iswidelyreferencedbycompaniesleadinginnovationinAI.Todiscussthis

report,our

publications,

or

ourservices,pleasegetintouchat

contact@artificialanalysis.ai.Commentary

?

TheperformancegapbetweenUS

andChinesefrontiermodelssince

thereleaseof

ChatGPTin2022has

remainedpersistent,butisasnarrownowasithaseverbeen?DeepSeek’sopenweightsR1(May2025)modelleads

theChinese

AIlabs,

whileOpenAI’so3leadsmodels

releasedbyUSAI

Labs?DeepSeek

andAlibabahaveprimarilydriven

theChinesefrontier,

whileadvancesin

theUSfrontierhavebeenprimarilydrivenby

OpenAIo1,OpenAIo1-preview,…"

OpenAIGPT-4o,OpenAIOpenAI14B,

Alibaba

Qwen1.5

Chat

Q

B,Alibaba

DeepSeek

Baichuan

4,BaichuanQwen3235B

A22B,

AlibabaSeed-Thinking-v1.5,ByteDanceDeepSeekR1(Jan.25)DeepSeek

V3Qwen2.5Plus,Alibabab7AwChina’sleadingAIlabsarenowcloser

thanever

toUSleaders,

withtheleaddecreasing

frommore

thana

year

toless

thanthreemonthsUS&China:FrontierLanguageModelIntelligence,Over

Time

ArtificialAnalysis

IntelligenceIndexincorporates7evaluations:MMLU-Pro,GPQADiamond,Humanity'sLastExam,

LiveCodeBench,SciCode,

AIME,MATH-500Gemini2.5Pro,Grok3mini

Googlereasoning(high),

xAIo3-mini(high),OpenAIo3,OpenAIDeepSeekR1(0528,May‘25)Note:Someresults

wereestimatedbasedoncompanyclaimsandcomparableresultsSource:

Artificial

Analysis

IntelligenceIndexQwen

ChatQwen

Chat

7B,110B,

Alibaba72B,

Alibaba

72B,

Alibaba

GLM-4-Plus,Artificial

Analys

isInte

lligence

IndexGPT-4,GPT-3.5Turbo,

OpenAI

Zhipu

AI?Artificial

AnalysisQwenChat

01.AIQwen1.5

Chat

Yi-Large,GPT-4Turbo,OpenAIStep

2,StepFunLLM67BLlama2Chat,

MetaLlama65B,

MetaGemma7B,

GoogleQwen1.5

Chat72B,Alibaba72B,Alibaba,'QwenChat14B,AlibabaDeepSeekR1(0528,May‘25)DeepSeekR1(Jan.25)

Llama3.1

NemotronUltra253B

v1,NVIDIALlama3.3Nemotron

Super

49B,NVIDIALlama3.370B,RekaFlash3Qwen272B,

DeepSeek-V2.5AlibabaQwen1.5Chat110B,Alibaba?Artificial

AnalysisCommentary

?

TheChineseopen

weights

frontier

surpassedtheUSinNovember2024

with

thereleaseofQwQ32BPreview(overtakingMeta’sLlama3.1

405B)?

TheopenweightsleadershipofChinese

AIlabsisreflectiveof

the

approachof

thetopChinese

AIlabstooftenreleasethe

weightsof

their

flagshipmodels.

Thiscontrastswith

thetopUS

AIlabs,whichgenerallydo

notreleasethe

weightsoftheirleading

models,e.g.,

OpenAI,

Anthropic

andGoogle?China,sDeepSeekR1(January2025)

was

thefirstopen

weightsreasoningmodel

tobecompetitive

withOpenAI,s

o1?DeepSeek,sR10528

(May2025)is

themostintelligentopenweights

modelcurrentlyavailableTheChineseopen

weights

frontiersurpassed

theUSinNovember2024withAlibaba’sreleaseofQwQ32BPreview,R1consolidated

thisleadUS&China:Open

WeightsFrontierLanguageModelIntelligence,Over

Time

ArtificialAnalysis

IntelligenceIndexincorporates7evaluations:MMLU-Pro,GPQADiamond,Humanity'sLastExam,

LiveCodeBench,SciCode,

AIME,MATH-500Note:Someresults

wereestimatedbasedoncompanyclaimsandcomparableresultsSource:

Artificial

Analysis

IntelligenceIndexInstruct

405B,

Instruct

72B,Meta

AlibabaGrok-1,xAIDeepSeek

V3,DeepSeek

QwQ32B-Preview,Alibaba

\Artificial

Analys

isIntelligence

IndexLlama

3.1

Qwen2.5Qwen3235BA22B,

AlibabaDeepSeekLLM67BQwen

ChatMetaLeadingChinese

AIlabsDeepSeekand

Alibabahavesteadilyreleasednewmodels,

withDeepSeek

takingtheleadinlate2024LeadingChinese

AILabs:LanguageModelIntelligence,Over

Time

ArtificialAnalysis

IntelligenceIndexincorporates7evaluations:MMLU-Pro,GPQADiamond,Humanity'sLastExam,

LiveCodeBench,SciCode,

AIME,MATH-500DeepSeekR1 (0528,May‘25)Qwen3235BA22BCommentary

?

AsofMay2025,DeepSeekR10528

(May2025)maintains

anintelligenceedgeover

Alibaba's

Qwen3235BA22Bas

theleading

model

fro

m

aChinese

AIlab?Both

companieshaveembracedopenweightsstrategies,supportingwidespreadadoptionof

theirmodels

domestically,as

wellasinternationally?Overthelasttwoyears,bothDeepSeek

andAlibabahavefrequentlyreleasedmodels

-releasingnewmodelsatmost~3monthslater

than

theirpriorreleaseNote:Someresults

wereestimatedbasedoncompanyclaimsandcomparableresultsSource:

Artificial

Analysis

IntelligenceIndex?Artificial

AnalysisCommentary

?DeepSeekleaped

over

xAI,MetaandAnthropic

tobe

tiedas

theworld,s#2

AIlab

andtheundisputed

openweightsleader

with

therelease

ofR1-0528?R1-0528,simpressiveintelligence

uplift

wasapost-trainingupdate

withnochange

to

theoriginalV3/R1architecture-itremainsa

large671Bmodel

with37Bactive

parameters?

Thishighlights

thecontinuallyincreasingimportanceofpost-

training,particularlyforreasoningmodels

trained

with

reinforcementlearning(RL)techniques6860+48463523DeepSeek’smodelshavequicklyincreasedinintelligencesincetheirfirstpubliclanguagemodelreleaseinNovember2023DeepSeek,sModelRelease

Timeline:

Artificial

AnalysisIntelligenceIndexOnlyhighestintelligencegeneral-purposeDeepSeekmodelsshown,

task-specificmodelsexcluded.

ArtificialAnalysisIntelligenceIndexincorporates7

evaluations:MMLU-Pro,GPQADiamond,Humanity'sLastExam,LiveCodeBench,SciCode,DeepSeek67BDeepSeek

V2DeepSeek

V2.5DeepSeek

V3DeepSeekR1DeepSeekR1-0528Source:

Artificial

Analysis

IntelligenceIndexNov.2023May2024Dec.2024Jan.2025May.2025Dec.2024Release

dateAIME,MATH-500.20Commentary

?OpenAIhasbeentheclearleader

ofthe

AIintelligence

frontier,but

itsleadhas

diminished

asother

labsincludingGoogle,

xAIandAnthropichavenarrowedthegap?

AsofMay2025,OpenAI,so3isthe

mostintelligentUSmodel,andthe

mostintelligentmodeloverallIn

theUS,

therearenowmultiplecontenders

for

AIleadership,

withOpenAInolongerinadominantpositionat

the

frontierLeadingUSAILabs:LanguageModelIntelligence,Over

Time

ArtificialAnalysis

IntelligenceIndexincorporates7evaluations:MMLU-Pro,GPQADiamond,Humanity'sLastExam,

LiveCodeBench,SciCode,

AIME,MATH-500Gemini

2.5Note:Someresults

wereestimatedbasedoncompanyclaimsandcomparableresultsSource:

Artificial

Analysis

IntelligenceIndex/AArtificial

AnalysisClaudeOpus4

(ExtendedThinking)Grok3

mini

reasoning(high)Llama

4

MaverickGo3ProThe

focusof

thisreport,deep

divetofollow.1.“Big

tech”players

ExistinglargetechnologyplayersinChinaAlibabaNYSE:BABA

(MktCap:

~$300B)Largeecommerceplayerandhyperscaler(Alibaba

Cloud),

largestshareholderofAntGroupByteDance

luul

Byte

DancePrivateParentcompanyofDouyin(TikTok),Toutiao,andDoubao,the

largest

AIchatbotinChinawith~60MmonthlyactiveusersHuaweiPrivateGlobal

telcoleader,hyperscaler,andleaderindomestic

semiconductor

manufacturing

forAIchips.TencentHKG:700(Mkt

Cap:

~$575B)ParentcompanyofRiotGamesand

WeChat,the'all-in-one

‘superapp’ofChina;hyperscalerwiththeir

TencentCloud

offeringBaiduNASDAQ:BIDU(MktCap:

~$30B)China’slargestsearchengine,

andoperatorof

WenxinYiyan,an

AIchatbot

with~13M

monthly

activeusers(MAU)2.AIstartups

Smallerfocussed

AIcompanies3.Other

companies

with

AI

ambitions OthertechnologyplayersXiaomiHKG:1810

(Mkt

Cap:~$165B)China’sleadingconsumerelectronicsbrand.Poached

LuoFuli,

DeepSeekresearcher,

torun

AIlabandrecentlyreleasedMiMo

-

7B,a

small

reasoning

model.KunlunTechSHE:

300418(MktCap:

~$5.5B)Beijing-basedinternetgroup

with>300mMAUs;

owneroftheOperabrowser;LaunchedSkyWork

AImodelsandaccelerators360Security(Qihoo

360)SHA:601360(Mkt

Cap:

~$10B)China'slargestproviderofInternetandmobilesecurity.Launched

theZhinaoseriesofmodelsunder

the360AIbrandiFlytekSHE:002230(MktCap:

~$15B)Leading

voice

AIcompany

inChinawith>14,000employees.

Launched

theSparkseriesofmodelsMeituanHKG:

3690

(Mkt

Cap:~$110B)China’s

topshoppingplatform

with>6

00mDAUs.Co-founder

Wang

Huiwenreturned

tolead

AIefforts

.InvestorinfrontierlabsDeepSeek

deep

seekLeadingChinese

AIlaboriginatingoutofHigh-Flyer,an

AI-focussedquantitativetrading

firm;bestknownforDeepSeekR1Moonshot

AI

ChinaAI

Tiger

with2MChinesecharactercontext

window

model;ownerofpopular

AIchatbotKimi

with~13MMAUZhipu

ChinaAI

Tigers

punoutfromaresearchprojectat

TsinghuaUniversityin2019;

Strongcommitmenttoopen

weightsmodelsStepFun

FirstChinese

AIstartup

todevelopa1Tparammodel;

founded

by

JiangDaxin(ex-Chief

Scientist,MicrosoftResearch

Asia)MiniMax

ChinaAI

TigerandpublisherofTalkie

AIapp(4th

mostdownloaded

app

in

the

US

in1H24)01.AI

FoundedbyLeeKai-Fu(formerheadofGoogleChina);

working

closelywith

Alibabatosetupa“industriallargemodellab”Baichuan

川i!熊ChinaAI

Tiger

withafocusonmedical

AImodelsTherearemanyChinese

AIplayers,and

theycategorizeinto3broadarchetypesNOT

EXHAUSTIVESource:PressSearchDescriptionLargeecommerceplayerandhyperscaler(AlibabaCloud),largestshareholderof

AntGroupParentcompanyof

RiotGames

andWeChat;hyperscalerwith

theirTencentCloudofferingChina’slargestsearchengine,

andoperatorof

Wenxin

Yiyan,

an

AIchatbotwith

~13Mmonthlyactiveusers(MAU)AI

Strategy?

Releasing

leading

open?

Developing

leading?

BuildingAIinfrastructure?Focussingondeveloping?Focussingondeveloping(high-level)weightsmodels(Qwen)proprietarymodelsinfromchips(Ascend910C),proprietarymodelsproprietarymodelsandtooling(ModelScope)language,

image,andvideosystems(CloudMatrix384)?Integratingmodelsinto?Integrating

modelsinto?Offeringmodel?Integratingmodels

across?Developing

proprietaryconsumer-facingchatbotconsumer-facingsearchinferencethroughmodalitiesintoconsumermodelson

theirown(Yuanbao),

and

WeChatproduct,andchatbotAlibabaCloud

toother

businessesfacingproducts(TikTok,

Toutiao,Doubao)hardware(PanguUltra)?Offeringmodelinference?Offeringmodelinference

throughBaiduCloudto(Wenxin

Yiyan,

formerly

ERNIEBot)?Offeringmodelinferenceandchiplease

throughotherbusinesses?Offeringmodelinferencethrough

VolcanoEngineHuaweiCloud

tootherthroughBaiduCloudto(ByteDance’scloudoffering)businessesotherbusinessesBestNon-Qwen3-235B-A22BDoubao

1.5ProPangu5.0Large1Hunyuan

TurboSERNIE4.5ReasonerLLMIntelligence:

47Intelligence:

48*Intelligence:

47*Intelligence:

46*BestReasoner

LLMQwen3-235B-A22B(Reasoning)Intelligence:62Seed

Thinking

v1.5Intelligence:62*-Hunyuan

T1Intelligence:59*ERNIE

X1Intelligence:60*ConsumerTongyiQianwenDoubaoCeliaYuanbao,

YuanqiWenxin

Yiyan,

Wenxin

YigeApps(~150MMAU)(~110MMAU)(~200MMAU)(~40MMAU)(~13MMAU)MarketCap(US$)~300BPrivatePrivate~$575B~30BLeadingChinese

“big

tech”

firmsareallpursuing

frontier-class

AImodels,andseveralhaveconsumer

AIapps

withover100MusersIn-depthDescriptionofChinese

“big

tech”

Firms1.Nopubliclypublished

benchmarksorinternational

APIavailable;*Estimatedbasedoncompanyclaimsandcomparableresults

where

available,not

yetindependentlybenchmarkedbyArtificial

Analysis

Source:

Artificial

Analysis

IntelligenceIndex,

Crunchbase,,PressSearchParentcompanyof

Douyin(TikTok),

Toutiao,andDoubao,the

largest

AI

chatbotinChina

with~60monthlyactiveusersGlobal

telcoleader,hyperscaler,andleaderindomestic

semiconductormanufacturing

forAIchips.ceep

seek

?Moonshot

Al

階躍星辰StepfunDescriptionLeadingChineseAI

laboriginatingout

ofHigh-Flyer,anAI-focussedquantitativetrading

firm;bestknown

for

DeepSeekR1China

AI

Tigerwith2M

Chinesecharactercontextwindowmodel;owner

ofpopular

AI

chatbot

Kimi

with~13MMAUChina

AI

Tigerspunout

fromaresearchprojectat

TsinghuaUniversityin

2019;strongcommitmentto

openweightsmodelsFirstChinese

AIstartup

todevelopa

1Tparammodel;foundedby

JiangDaxin(ex-ChiefScientist,Microsoft

ResearchAsia)China

AI

Tigerandpublisherof

Talkie

AI

app

(4th

mostdownloadedappin

theUSin1H24)Founded

by

LeeKai-

Fu

(ex-Head

ofGoogleChina);workingcloselywith

Alibabaonan“industriallargemodellab”China

AI

Tigerwitha

focusonmedical

AImodelsFoundedMay2023March2023June2019April2023December

2021March2023April2023NotableinvestorsTencent騰訊Tencent騰訊Tencent騰訊Tencent騰訊SSVXiaohongshu

(Red

note)BestNon-DeepSeek

V3Moonshot

v1GLM-4-32BStep-2MiniMax-Text-01Yi-LightningBaichuan

4ReasonerLLMIntelligence:52Intelligence:30*Intelligence:

46Intelligence:

44Intelligence:

40Intelligence:37*Intelligence:31*BestReasoner

LLMDeepSeekR1Intelligence:68KimiK1.5Intelligence:58*GLM-Z1-32BIntelligence:56Step-R1-V-MiniIntelligence:55*--BaichuanM1Intelligence:52*ConsumerDeepSeekChatKimiChatGLMYuewen,PopDuckTalkie

AIYiChatBai

XiaoyingApps(~100MMAU)(~25MMAU)(~9MMAU)(<1MMAU)(~35MMAU)(<1MMAU)(<1MMAU)Thereisa

deepecosystemof

well-fundedChinese

AIlabstartups

withcompetitive

AImodels,manyhave

financialbackingbyChinese“big

tech”

playersIn-depthDescriptionofChineseStartups*Estimatedbasedoncompanyclaimsandcomparableresults

where

available,not

yetindependentlybenchmarkedbyArtificial

Analysis

Source:

Artificial

Analysis

IntelligenceIndex,

Crunchbase,,PressSearchFundingraised

(US$)~0.85BUndisclosedUndisclosed~0.2B~1.0B~1.6B~1.1BWhileUS

AIlabsmaintainanoverallleadin

theintelligence

frontier,Chinese

AIlabsarenolonger

farbehindChineseandUSLabs,LeadingLanguageModels(non-exhaustive)*Estimatedbasedoncompanyclaimsandcomparableresults

where

available,not

yetindependentlybenchmarkedbyArtificial

Analysis

Source:

Artificial

Analysis

IntelligenceIndexArtificialAnalysis

IntelligenceIndexincorporates7evaluations:MMLU-Pro,GPQADiamond,Humanity'sLastExam,LiveCodeBench,SciCode,

AIME,MATH-50051

vlli

40?Artificial

Analysis

USA

ChinaQwen3235BA22B(Reasoning),Ali

babaSeed-Thinking-v1.5,ByteDance65

62DeepSeekR1

(May2025),DeepSeekClaudeOpus

4Thinking,AnthropicBaichuanM1(Preview),

BaichuanGrok

3MiniReasoning

(high),xAI56

55*Kimi

K1

.5

Preview

,

MoonShotLlama4Maverick,MetaYi-Lightning,01

AIstep-r1-

v-mini,

StepFuno4-mini

(high),

OpenAIMiniMax-Text-01,

MiniMaxGemini2.5

Pro,

Googleglm-z1-

32b,

ZhipuG68Hunyuan

T1,

TencentERNIEX1,BaiduMiMo-7B,

Xiaomi59

*60

*58

*62*52*37*546770686862

61

60

5952

51

50

48

4734Non-reasoning:Open

WeightsLanguageModels,byCountry

ArtificialAnalysis

IntelligenceIndex,leadingopen

weightsnon-reasoningmodelsm

USAmCanada

ChinaFrance53

5147

4641

40

40

40

38

38Today'sopen

weights

frontierisledbyDeepSeek(bothreasoningandnon-reasoningmodels)Reasoning:Open

WeightsLanguageModels,byCountry

ArtificialAnalysis

IntelligenceIndex,leadingopen

weightsreasoningmodelsmUSA

m

China

DeepSeekV3(Mar,25)Llama4Qwen3GLM4-32BLlama3.3Maverick

235BA22BInstruct

70BCommandAPhi-4

MiniMax-Text-01Gemma327BInstructMistralLarge2(Nov,24)DeepSeekQwen332B

DeepSeek

R1(Reasoning)R1

Distill(JanuaryQwen32B2025)Llama3.3

Phi-4Nemotron

(reasoning)

Super

49BReasoningSource:

Artificial

Analysis

IntelligenceIndexLlama3.1

Nemotron

Ultra253B

ReasoningQwen3235B

A22B

(Reasoning)DeepSeek

R1

(May

2025)DeepSeek

R1

DistillLlama70BReka

Flash

3MiMo7B

RLIncludesmodels

releasedsinceJanuary2024,NotexhaustiveCompanyLanguage,ReasoningLanguage,Non-ReasoningText

toSpeechSpeech

toSpeechImageGen.

VideoGen.3D

Gen.AlibabaTencentQwen3235BA23BQwen2.5OmniWan

2.1LHMERNIE

X1ERNIE4.5-

-I-RAGUniVGXilingSeed-Thinking-v1.5Doubao-1.5-proSeed-TTS

Doubao-1.5-proSeedream3.0Seaweed-7BImageDreamPanGu

UltraMoEPangu5.0-

-PixArt--Hunyuan-T1Hunyuan-TurboS-

-HunyuanDiT

HunyuanVideoHunyuan3DAI

StartupsDeepSeekBaichuanDeepSeekR1DeepSeek

V3-

-JanusPro7B

--Kimik1.5Moonshot-v1-

Kimi-Audio-

--Gl

m-z1-32bGlm-4-32b-

-CogView4

CogVideo-Step-R1-V-MiniStep-2Step-audioStep-1x

StepVideoT2VStep1X3D-

MiniMax-Text-01Speech-02HD

-Image-01Video-01--

Yi-Lightning-

--

--Baichuan-M1Baichuan

4Baichuan-Audio-

--Chinesecompaniesplayacrosseverymajormodern

AImodality-includinglanguagemodelsandall

typesofmedia

generationmodelsLeadingmodelforeachmodality,bycompany

NoModelResearchModelOnlyPubliclyReleased

ModelMoonShot

AIZhipuStepfunMiniMax01.AIBaiduByteDanceHuawei“Big

tech”

PlayersSource:PressSearchLeadingText

toImageModels,Artificial

AnalysisArenaELOsCommentaryText

toImagemodelELOs

from

theArtificial

AnalysisImageArena

Canada

Germany?Chinaand

theUSareateffectiveparityin

text

toimagegeneration,

withOpenAI’sGPT-4omaintaininga

4

pointELOleadoverByteDance’s1,165

1,1611,111

1,1091,0941,054Seedream3.0?

TexttoImageleadershail

fromawidediversityof

geographies,unlikelanguagemodels

wheretheUSandChinadominate,drivenbysmallerstartupswithaspecificfocusonimagemodels?Someoldermodelscontinue

to

performwellagainstrecentreleases,indicatingaslowerinnovation

ratethaninlanguage

models,

withRecraftV3(October2024)remainingnotably

competitive

despitenewerentrantsText

toImagemodelqualitybetweenUSandChinareachedparityin

2025ReleaseMay,April,October,April,December,March,March,October,February,July,Date20252025202420252024202520252024202520241,087

1,085

1,082Source:

Artificial

Analysis

MediaGeneration

ArenaLeaderboardsAArtificial

AnalysisMidjourney

v6.1,MidjourneyIdeogram

3.0,IdeogramSeedream

3.0,Bytedance

USA

ChinaFLUX.1.1

[pro],

BFLImagen

4

Ultra

Exp.Image-01,

MiniMaxRecraft

V3,RecraftHiDream

-I1-DevGPT-4o,

OpenAIReve

Image1,047Commentary

?

TheUSmaintainsaleadin

textto

videomodels,

withGoogle's

Veo3

PreviewachievingasignificantlyhigherELOin

theArtificial

Analysis

Video

ArenathanKuaishou'sKling2.0?Chinesecompanies

are

widely

competitivein

the

text

tovideo

space,

withlabssuchasMiniMax

andAlibabaperformingatcomparablelevels

toOpenAIand

PikaText

toVideoisledbyUSandChineselabs,

withGoogle’s

Veo3leading

thecategorybya

widemarginLeadingText

toVideoModels,

Artificial

Analysis

Video

ArenaELOs

Text

to

VideomodelELOs

from

the

Artificial

Analysis

Video

ArenaReleaseDecember,

April,December,September,December,February,December,

October,June,

June,Date

2024

20252024202420242025202420242024

20241,247Veo3

Preview,

Google1,113Kling2.0,

Kuaishou967Ray

1,

LumaLabs1,005HunyuanVideo,Tencent1,027Wan

2.1,

Alibaba1,037Pika

2.0,Pika

Art1,039T2V-01,

MiniMax1,053Sora,

OpenAI1,000Mochi1,

Genmo989Gen

3

Alpha,

RunwaySource:

Artificial

Analysis

MediaGeneration

ArenaLeaderboardsAArtificial

Analysis

USA

ChinaCommentary

?Google,sVe