摩根士丹利-中國AI基礎(chǔ)模型（亞太）：價(jià)格戰(zhàn)之上的智能戰(zhàn)-China AI Foundation Models Asia Pacific Intelligence War Over Price War-20260809

Morganstanley

August9,202610:03PMGMT

RESEARCHIdEA

ChinaAIFoundationModels|AsiaPacific

IntelligenceWarOverPriceWar

WHAT’sCHANGED

From

HK$990.00

To

HK$1,700.00

Z.AICO.,LTD.(2513.HK)PriceTarget

MiniMax(0100.HK)

PriceTargetHK$1,100.00HK$900.00

WeviewChina’sLLMindustryasbuildingahealthier

commercializationenvironment—asopposedtothemarketconsensusthatChina’sopen-weightmodelswilldrive

commoditizationandapricewar.

WeseeChinaLLMsacceleratingcommercializationinapositiveway,frompricecompetitiontointelligence-drivenmonetization.Threeshiftsunderpinourview.

First,pricingisturningmorerational,withDeepSeek’slatestAPIpriceincrease

signalingimprovingindustrypricingdiscipline.Second,open-weightmonetizationistighteningasvendorsmovetowardmorerestrictivelicensesandexpandfrom1PAPIsalesto3Prevenuesharing.Third,modelscaleissteppingup,withKimiK3’s2.8T

andQwen3.8-Max's2.4TparametersmarkingChina’sentryintothelarge-parameterera.Asfrontiermodelsscale,risingrequirementsforcapital,compute,infrastructureoptimizationandengineeringexecutionshouldwidencompetitivemoats,favoring

well-resourcedleadersandacceleratingindustryconsolidation.Together,webelievethesetrendspointtobettermonetization,higherbarrierstoentryandfurther

industryconsolidation.

Intelligence,notprice,remainstheultimatemoat:Frontierplayerscanmove

down-marketwithcheapermodels;average-modelvendorsfaceamuchharderpathmovinguptoSOTA.Astrainingandcomputerequirementsrise,webelievewinnerswillincreasinglybethosethatcansustaintheflywheelofbettermodels→strongermonetization→morefundingandcompute→greatermodelinvestment.

CommercializationandAGIarethereforecomplementary,inourview:bettermonetizationfundsintelligence,whilebetterintelligencedrivesmonetization.

Stockviews:Withinourcoverage,weseethisflywheelformingfaster

atZ.ai:GLM-

5.2

,

improvingcomputeaccessandfreshfinancingsupportstrongergrowth

,

driving

our2026ARRestimatetoUS$2bnandPTtoHK$1

,

700

.WeremainconstructiveonMiniMax,withtheupcomingM3updateand2-3TM3Proaskeycatalysts,butseeamoreback-end-loadedgrowthtrajectory;wemaintainourUS$1bn2026ARR

forecastandlowerourPTtoHK$900.AlthoughwearepositiveinChinaLLMs'

competitivedynamics,welowerourbearcasemultiplestoreflectpossibilitiesofintensifiedcompetitionwherebyglobalpeerslaunchbettermodelswithlower

prices.Wealsoraise2027eR&DexpensesaboveUS$1bnforbothcompaniesto

reflectmuchhigherinvestmentneededfortraininglargeparametermodels.WearealsobullishonAlibabaforitsfullAIstackandadvantagesincompute,with

multiyearcloudgrowthandintactmarginexpansion.

MoRGANSTANLEyAsiALimiTED+

GaryYu

EquityAnalyst

Gary.Yu@

+8522848-6918

LydiaLin

EquityAnalyst

Lydia.Lin@

+8522239-1572

GREATERCH?NA?TSERv?cEsANDSoFTwARE

AsiaPacific

IndustryView?n.Line

Relatedreading:GlobalThematics:WeighingIn:Open-WeightsModels&3StatesoftheWorld(3Aug2026)

MorganStanleydoesandseekstodobusinesswith

companiescoveredinMorganStanleyResearch.Asaresult,investorsshouldbeawarethatthefirmmayhaveaconflictofinterestthatcouldaffecttheobjectivityofMorganStanley

Research.InvestorsshouldconsiderMorganStanley

Researchasonlyasinglefactorinmakingtheirinvestmentdecision.

Foranalystcertificationandotherimportantdisclosures,refertotheDisclosureSection,locatedattheendofthisreport.

+=Analystsemployedbynon-U.S.affiliatesarenotregisteredwithFINRA,maynotbeassociatedpersonsofthememberandmaynotbesubjecttoFINRArestrictionson

communicationswithasubjectcompany,publicappearancesandtradingsecuritiesheldbyaresearchanalystaccount.

2

Morganstanley

RESEARCH

IDEA

InvestmentThesis

WearebullishonChinaLLMpricingtrends.On8August,DeepSeekannounceda

significantAPIpriceincrease,markinganotableshiftforavendorthathashistorically

combinedcompetitiveperformancewithindustry-lowpricing.Weseethreelikelydrivers:

1)strongdemandforV4modelssupportinggreaterpricingpower;2)theneedtobalancemarket-sharegainswithhealthygrossmarginstofundcontinuedfrontier-model

investment;and3)potentiallyhigherinferencecostsfromgreateruseofdomestic

chipsets.Weseethisasapositivesignalforindustrypricingdiscipline:pricewarsare

unlikelytobeasustainablecompetitivestrategy.Asmodelintelligenceandscalecontinuetoimprove,weexpectpricingfornewChineseLLMstotrendhigher,withcompetition

increasinglyshiftingfrompricetocapabilityandcapacity.

Moreimportantly,webelievemodelintelligence—notprice—isthekeydeterminantoflong-termcompetitivepositioninginLLMs.AsdiscussedinMoreBangForTheBuck(27Apr2026),weexpecttheLLMmarkettobifurcateintoahigh-endsegment,differentiatedbyfrontier-levelintelligence,andalow-endsegment,wheremodelcapabilitiesare

increasinglycomparableandpricecompetitionismoreintense.However,wedonotbelievethesetwosegmentswillnecessarilybeservedbydifferentvendors.Frontiermodeldeveloperscanrelativelyeasilyofferlower-cost,lower-capabilityvariantstoaddressthemassmarket,whilethereverse—movingfromanaveragemodeltothefrontier—isconsiderablymoredifficult.

Thisasymmetrymattersbecausesustainingfrontierintelligencerequirescontinuousandsubstantialinvestmentinmodeltraining.Inourview,pursuingmarketshareprimarily

throughaggressivepricingoflower-tiermodelscouldthereforebeariskystrategyatthisstageoftheindustry.Lowergrossmarginsmayconstraininternallygeneratedfundingforfrontiermodeldevelopment,whileaweakertechnologypositioncouldalsomakeexternalfinancingmoredifficulttosecure.WethereforeviewcommercializationandthepursuitofAGIascomplementaryratherthanseparateobjectives:commercializationprovidesthefinancialresourcesrequiredtosustaininvestmentatthefrontier,whilefrontier

capabilitiesunderpinlong-termmonetizationandcompetitivedifferentiation.

Exhibit1:ChinafrontierLLMcomparison

Z.ai

GLM-5.2

DeepSeek

V4-Flash

Alibaba

Qwen3.8-Max

Moonshot

K3

AIlab

Model

Xiaomi

MiMoV2.5Pro

MiniMax

M3

License

ReleasedateAug-26Jun-26Jun-26Jul-26Jul-26Apr-26

OpenWeight

OpenWeight

OpenWeight

OpenWeight

OpenWeight

OpenWeight

Totalparameter(B)

428

754

2800

284

1000

2400

Contextwindow(K)

1000

1000

1000

1000

1000

1000

APIprice(Rmb/mntokens)

Input(uncached)

Input(cachedread)

3.150.63

8.002.00

20.002.00

1.000.02

3.000.03

12.001.50

Output

36.00

12.60

28.00

100.00

2.00

6.00

AAIntelligenceIndex

58

45

53

60

52

43

Source:Companydata,MorganStanleyResearch

Exhibit2:ChinafrontierLLMpricevs.intelligence

30

020406080100120

Outputprice(Rmb/mntoken)

K3

Qwen3.8-Max

GLM-5.2

V4-Flash

M3

MiMoV2.5Pro

65

60

55

50

45

40

35

ArtificialAnalysisIntelligenceIndex

Source:ArtificialAnalysis,companydata,MorganStanleyResearch

Morganstanley

RESEARCH

IDEA

Exhibit3:ChinaLLMsaverageAPIpricing

Input(Rmb/mntoken)Output(Rmb/mntoken)

25.0

21.9

20.018.118.1

12.2

15.013.412.7

10.0

4.9

5.03.32.22.84.03.7

0.0

1Q252Q253Q254Q251Q262Q26

Source:Companyofficiallistingprice(incl.Bytedance,Alibaba,Baidu,Tencent,MiniMax,

Z.ai

,MoonshotandDeepSeek);MorganStanleyResearch

Exhibit4:ChinaLLMsaverageAPIpricingas%ofUSpeers

8%

6%

3Q254Q251Q262Q26

InputOutput

19%18%19%

17%

15%

14%

20%

18%

16%

14%

12%

10%

8%

6%

4%

2%

0%

7%

6%

2Q25

6%

5%

1Q25

Source:Companyofficiallistingprice(incl.Bytedance,Alibaba,Baidu,Tencent,MiniMax,

Z.ai

,MoonshotandDeepSeek);MorganStanleyResearch

Commercializationaccelerating:tighteropen-weightlicensesandashiftfrom1Pto3Pmonetization.ChineseLLMshavehistoricallybeenassociatedwithpermissiveopen-

weightmodels,withApache2.0/MIT(L1)licensesallowingbroadthird-partydeploymentandcommercialization.Asmodelcapabilitiesimprove,however,weseevendors

increasinglymigratingtowardmorerestrictiveL2/L3licensesthatretainopenweightswhilelimitingcommercialusage.Forexample,Moonshot’sK3licenserequiresMaaS

providersgenerating>US$20mninannualrevenuefrommodeldistributiontonegotiateseparatecommercialterms,whileReuters(Aug7)reportedthatAlibabamayintroducerevenue-sharingrequirementsforlarge-scaleusersoffutureopen-weightQwenmodels.

WeseethisasanotherpositivemonetizationsignalfortheChinaLLMindustry.TighterlicensescouldextendmonetizationbeyonddirectAPIsales(1P)towardrevenuesharingwithCSPsandmodelaggregators(3P),expandingtheaddressableARRpoolbeyondthecomputecapacityownedorrentedbyLLMvendorsthemselves.

Exhibit5:AIfoundationmodellicensecategories

Level

LicenseCategory

Definition

Weights

Available

CommercialUse

Fine-tuning/

Derivatives

Redistribution

MaaS/API

Commercialization

KeyRestrictions

L1

PermissiveOpenSource

Modelweightsarereleasedunderstandard

permissivelicensessuchasApache2.0orMIT,

allowingunrestrictedcommercialuse,

modificationandredistributionsubjecttolimitedlicenserequirements.

Yes

Yes

Yes

Yes

Yes

Minimalrestrictions;typicallyrequiresretentionof

copyright/licensenotices;Apache2.0alsoincludesexplicitpatent

provisions

L2

CommercialOpen-weight

Modelweightsarepubliclyavailableand

commercialuseisgenerallypermitted,but

customlicensetermsmayimposerevenue,scale,attribution,redistributionorMaaS-related

restrictions.

Yes

Yes,withconditions

Yes

Yes,withconditions

Yes,withconditions

Mayincluderevenue,user-scale,

attribution,redistribution,usage,orMaaSrestrictions

L3

Non-commercialOpen-weight

Modelweightsarepubliclyavailableforresearchornon-commercialuse,whilecommercial

deploymentrequiresseparateauthorizationoracommerciallicense.

Yes

No,unlessseparately

licensed

Yes,forpermitted

non-commercial

purposes

Usuallylimited

No,unlessseparately

licensed

Commercialuseisprohibitedor

requiresseparateauthorizationfromthemodelprovider

L4

Closed/API-only

Modelweightsarenotpubliclyavailable;thirdpartiesaccessthemodelprimarilythroughfirst-partyorauthorizedAPIsandproducts.

No

Yes,throughapproved

services/APIs

Generallyno

No

Yes,throughproviderAPI

Noaccesstomodelweights;usagegovernedbyproviderterms,API

policiesandpricing

Source:MorganStanleyResearch

MoRGANSTANLEyREsEARcH3

4

Morganstanley

RESEARCH

IDEA

Exhibit6:Majoropen-weightLLMlicenses

Company

Model

License

Level

CommercialUse

KeyRestriction

DeepSeek

DeepSeek-V3.2

MIT

L1

Yes

NonebeyondstandardMITrequirements

DeepSeek

DeepSeek-V4-Flash

MIT

L1

Yes

NonebeyondstandardMITrequirements

DeepSeek

DeepSeek-V4-Pro

MIT

L1

Yes

NonebeyondstandardMITrequirements

Z.ai

GLM-5.1

MIT

L1

Yes

NonebeyondstandardMITrequirements

Z.ai

GLM-5.2

MIT

L1

Yes

NonebeyondstandardMITrequirements

Moonshot

KimiK2.6

ModifiedMIT

L2

Yes

Brandingrequiredabove100mMAUorUS$20mmonthlyrevenue

Moonshot

KimiK3

KimiK3License

L2

Yes,withconditions

LargeMaaSoperators(>US$20mgrouprevenue/12months)requireseparateagreement;brandingrequiredabovescalethresholds

MiniMax

MiniMaxM2.7

Non-commercialLicense

L3

No,unlessauthorized

Allcommercialuserequirespriorwrittenauthorization

MiniMax

MiniMaxM3

MiniMaxCommunityLicense

L2

Yes,withconditions

Commercialattribution+notice;>US$20mannualproduct/servicerevenuerequiresauthorization

Alibaba

Qwen3.8-Max

TBD

TBD

TBD

Open-weightreleaseannounced;officiallicensenotyetverified

Source:Companydata,MorganStanleyResearch

Viewsonopen-weightmodels:moreadoption,notlesscompute.Webelieveinvestorconcernsthatcheaper,moreefficientopen-weightmodelscouldstructurallyreduceAI

infrastructuredemandareoverstated.Instead,lowerinferencecostsshoulddrivehigherusagethroughJevonsParadox,acceleratingAIdiffusion.Adoptionisalreadymeaningful—63%ofsurveyedenterprisesuseopenmodelsalongsideclosedmodels—though

deploymentsremainconcentratedinusecasesrequiringhighfrequency,customizationordatacontrol.

Openweightsoffercompellingeconomics,customizationanddeploymentflexibility,butarenot“free”:hosting,fine-tuning,engineering,securityandliabilityremainmeaningfulcosts.Wethereforeexpectopenandclosedmodelstocoexistratherthanconvergetoawinner-takes-allmarket,drivinganincreasinglymulti-modelenterprisestackandcreatingincrementalvalueforrouting,orchestration,observability,governanceandsecurity.

(PleaserefertoGlobalThematics:WeighingIn:Open-WeightsModels&3StatesoftheWorld(3Aug2026)formoredetailsonopen-weightmodels).

ChineseLLMsareenteringthelarge-parameterera—raisingthecompetitivebar.

Through1H26,Chineseopen-weightLLMsgenerallyremainedbelow1Ttotalparameters.ThisbegantochangeinJuly2026,whenKimiK3launchedwith2.8Tparameters,

becomingthefirstmajorChineseopen-weightmodeltocrossthe2Tthreshold.Itsstrongperformancegloballyprovidesfurtherevidencethatscalingremainsaneffectivepathtoimprovingmodelintelligence.Weexpect2-3Tparameterstoincreasinglybecomethe

baselineforfrontierChinesemodelsin2H26,withseverallarge-scalelaunchespotentiallyahead,includingMiniMaxM3ProinSep-Oct,Zhipu’snext-generationGLMmodelin

October,andpotentiallyAlibaba’sQwen4.Mediareports(LatePost,Aug6)alsosuggestByteDanceisexploringa5T+parametermodel.

MoRGANSTANLEyREsEARcH5

Morganstanley

RESEARCH

IDEA

Exhibit7:UpcomingLLMpipeline

Source:Companydata,MorganStanleyResearch

Biggermodels,higherbarriers:ContrarytotheviewthatLLMsfacelimitedbarrierstoentry,webelievethetransitiontowardlargermodels—particularlythosewithhigher

activeparametercounts—shouldwidenthecompetitivemoat.Scalingrequiresnotonlygreatertraininginvestmentandcompute,butalsostrongercapabilitiesininference

efficiencyandinfrastructureoptimization.Inourview,thisincreasinglytestsvendors

acrossmodelknow-howandarchitecturaljudgment,systemsengineering,organizationalexecution,financingandcomputesourcing.Asfrontiermodelsizescontinuetoscale,thecapitalandexecutionrequirementsshouldriseaccordingly,favoringwell-resourced

leadersanddrivingfurtherconsolidationoftheLLMmarket.

Exhibit8:GlobalSOTAmodelintelligenceindexranking

70

60

50

40

30

20

10

0

KimiK3

(max)

DeepSeekV4Flash0731(max)

MiniMax-M3

Qwen3.8Max

US

China

GLM-5.2(max)

Source:ArtificialAnalysis,MorganStanleyResearch

Exhibit9:ChinafrontierLLMsizevs.intelligence

K3Qwen3.8-Max

GLM-5.2

MiMoV2.5Pro

10001500200025003000

Totalparameters

V4-Flash

M3

0500

ArtificialAnalysisIntelligenceIndex

65

60

55

50

45

40

35

30

Source:ArtificialAnalysis,companydata,MorganStanleyResearch

6

Morganstanley

RESEARCH

IDEA

Z.ai:

FormingtheFlywheel

Whathasbeenpricedin:fromChinaNo.1toTier

1.Z.ai

sharesrosec.13xfromitsJanuary2026IPOthroughearlyJuly,significantlyoutperformingtheHSTECHIndex’sc.14%returnoverthesameperiod.However,followingKimiK3’slaunchon17July,thestockcorrectedby>50%atitstrough.Weattributemuchofthisvolatilitytoaresetinmarket

expectations

aroundZ.ai

’scompetitivepositioning:fromperceivedChinaNo.1followingGLM-5.2’slaunch,backtoaTier-1playerafterthelarger-scaleK3surpassedGLM-5.2onkeyperformancebenchmarks.

Webelievethemarketmaybeover-emphasizingpoint-in-timemodelrankings.TheNo.1positionhashistoricallyrotatedevenamongleadingUSmodeldevelopers,andwewouldexpectsimilar—ifnotgreater—rotationinChinaasmodelreleasecycles

accelerate.Inourview,themoreimportantquestionis

notwhetherZ.ai

holdstheNo.1positionatanygivenpointintime,butwhetheritcansustainapositivedevelopmentflywheel:leadingmodelcapabilities→strongeruseradoptionandmonetization→greaterreinvestmentintotrainingandcomputetogetherwithenhancedfinancingchannels→continuedmodelimprovement.Weseetheabilitytosustainthiscycle,ratherthantemporarybenchmarkleadership,asthemoredurabledeterminantoflong-termcompetitivepositioning.

We

seeZ.ai

formingsuchaflywheel:

1.World-classmodelperformance:GLM-5.2hasreachedglobaltopLLMs'

performancelevelnotjustinthebenchmarksbutalsoinglobaldevelopers'

communityfeedbacks,

boostingZ.ai

fromChina'stotopLLMplayertoaleadingglobalplayer.OnJune16,

Z.ai

'sChiefScientist,JieTang,statedonXthatthe

companywouldreleaseamodelmatchingthecurrentbest-performingLLM

developedbytheUSbefore1Q27,pointingtoarapiditerationtrajectoryontopofGLM-5.2.We

expectZ.ai

toreachthistargetearlier,inOct2026,withitslarge

parameterversionandreturntotheSOTAposition.

2.Improvingcomputesupply:Inferencecomputehasalwaysbeenaconstraintto

Z.ai

'sARRgrowthandespeciallyforoverseas

compute.RecentlyZ.ai

hasbeeninnegotiationswithAWS,andGLM-5.2APIhasbeenintegratedintoAWS

Marketplace.Weexpectsuch3PbusinessmodeltofurtherexpandtootheroverseasCSPsandAPIplatforms,tacklingoverseascomputeconstraints.

3.Positivefinancingcycles:OnJuly13,

Z.ai

completedanH-shareequityplacementbyraisingHK$31.4bn(US$4bn)with4.25%dilution.Thecompanyaimstoallocate55%(US$2.2bn)ofthefundingtomodeltrainingandcomputeexpansion.

Meanwhile,apotentialA-sharelistingcouldprovideanotherRmb15bninfunding.Withmorefunding,thecompanycouldgeneratemorerevenuefromexpanded

computeandfurtherreinvestintomodeltraining,formingapositivedevelopmentcycle.

Downsiderisk-willChinesegovernmentscurbAImodelexports?ReutersonJuly7

reportedthattheChinesegovernmentislookingtocurboverseasaccesstoChina'stopAImodelswithBytedance,Alibaba

andZ.ai

specified.Wedon'tthinkthisismaterialfor

multiplereasons.TheChinesegovernmenthasalwaysencouragedexportsofhightech,

MoRGANSTANLEyREsEARcH7

Morganstanley

RESEARCH

IDEA

includingAI.Datasecurityissuesexistinout-reachingaccesstoforeignmodelsratherthanincomingaccesstodomesticmodels.MostChinesemodelsareopen-weight,and

technicallyopen-weightmodelscan’tbe"curbed"foraccess.OnlyBytedance,Alibabaand

Z.ai

werementioned,andnotDeepSeekorMoonshot,whicharealsoconsideredcurrenttieroneplayersinChina.However,webelievethegreaterdownsideriskisfromtheUSgovernmentside.

Z.ai

'sfutureroadmap:

InZ.ai

'sinternalletterreleasedonJuly11,JieTangpointedoutthecompany'sfocusonAGIwithnextfewtechnologymilestonestocrossfromlong-horizontasks,autonomousagentsystemtoself-evolving.Withthesetargets,

Z.ai

releasedthe

"TouchHigh"plantoprioritizeAGItargetsovershort-termcommercialization,strategicallyinvestinginlong-horizontasks,autonomousagentsystem,fullyselftrainingandextremesecuritygovernance,atthesametimeremainingopen-weighttosupportecosystem

development.

ARRandvaluation:Z.ai

disclosedthatitsARRalreadyhititsfull-yeartargetofUS$1bnin1H26.We

expectZ.ai

'sARRgrowthtoaccelerateafterGLM5.2'slaunch,togetherwith

inferencecomputesupplyexpansionin2H26.Weraiseour2026ARRestimatefromUS$1bntoUS$2bn.Our2026bullandbearcaseARRestimatesareUS$3bnandUS$1.5bn,respectively.Ourbullbase

reflectsZ.ai

'sgreaterthanexpectedcomputeexpansion,

especiallyoverseas,whileourbearcaseincorporatesintensifiedcompetitionwherebycompetitorslaunchbettermodelswithlowerprices.

Exhibit10:

Z.ai

modelpricetrend

28

24

20

1313

33

1.5

30

25

20

15

10

5

0

7

5

GLM-4.7(Dec

2025)

GLM-4.6(Sep

2025)

GLM-5.0(Feb

2026)

GLM-5.1(Apr

2026)

GLM-5.2(Jun

2026)

GLM-4.5(Jul

2025)

BlendedinputOutput

Rmb/mntokens

8

6

Source:Companydata,MorganStanleyResearch,Inputpriceblendedwithallcontextlength

Exhibit11:ZaiARRestimates

US$1.5bn

GLM-5.2

GLM-4.7

GLM-5

Jun-25Sep-25Dec-25Mar-26Jun-26Sep-26Dec-26

BaseBearBullUS$3bn

1,000

500

-

GLM-4.5GLM-4.6

2,500

2,000

US$2bn

US$mn

3,000

1,500

.

Source:MorganStanleyResearchestimates

8

Morganstanley

RESEARCH

IDEA

Exhibit12:

Z.ai

three-stepAIdevelopmentroadmap

Source:Companydata,MorganStanleyResearch

MoRGANSTANLEyREsEARcH9

Morganstanley

RESEARCH

IDEA

MiniMax:WhyWeRemainOptimistic

Whathasbeenpricedin:disappointmentinM3releaseandlarge-volumeunlock.

MiniMax'ssharepricerose38%sinceIPO,

significantlyunderperformingZ.ai

especially

afterM3wasreleasedonJun1.DespiteM3alsoachievinghighranking,itbenchmarkswithgoodagenticandfull-modalityunderstandingcapabilities;itssmallsize(428Btotal

parametersand23activeparameters),doubledpricevs.M2.7andrelativeshortagein

codinghastriggerednegativefeedbackfromdevelopercommunities,leadingtoadropinsharepricetothetroughby77%sinceJun1.Inaddition,MiniMax'scornerstoneandpre-IPOinvestorunlockonJul8released10xfreefloat(from5.44%oftotalsharesto54.38%oftotalshares),leadingtogreatersellingpressure.

However,wearenotasnegativeonMiniMax:

1.MultiplenewLLMupgradestobereleased:WeviewM3moreasan

experimentalLLMtotestnewarchitecture(MiniMaxSparseAttention)before

scalinguptolarge-sizemodels.Theexperimenthasbeensuccessfulintermsof

inferenceefficiencyimprovementinalongercontextlengthat1Mtokens,andweexpecttheM3seriesmodeltogeneratehighergrossmarginthanM2series

despiteAPIpriceendingupthesamewiththeM2series.AlthoughthecurrentM3islessthanafullversionmodel,weexpectthecompanytoreleaseanupgradeinJul-Augwithprogressivepost-trainingtoenhancecodingcapability.MiniMax

announcedthattheM3Prowith2.7TtotalparameterswillcomeoutinSep-Oct,whichshoulddeliveradvancedperformancebeingthelargestChineseLLMsofar,aswellasdemonstratingpricingpower.

2.LLMismorethancoding:WebelievecodingisthefirstAIscenariothathas

crossedthetechnologytippingpointandcommercializationinflectionpoint.AndcodingcapabilityisoneofthemostimportantcapabilitiesofLLMsduetoits

stronggeneralizationfeatureaswellasbeingthefoundationofmodelself-

evolving.However,webelieveLLMwillpenetrateintomoreproductivity

scenarios,expandingfromtheprogrammermarkettothebroaderknowledgeworkermarketincludinglaw,finance,consultingandhealthcare,whichcan'tbesimplysatisfiedwithcodingcapability.MiniMaxhasbeenfocusingonverticalindustrytrainingdataandits10xTeaminitiative(kick-startedinMay2026)hasbeenrecruitingexpertsfromvariousindustries.ThecompanycouldpotentiallystandoutwhenAIcompetitionshiftingfromcodingtootherverticals.

3.Multi-modalityisunderappreciated:Capitalmarkets'attentionappearsfully

allocatedtoLLMsratherthanmulti-modalmodels.AndMiniMaxhassuspendeditsmulti-modalmodelreleasessinceHailuo2.3inOct2025toconcentrate

investmentinnextgenerationdevelopment,whichhascreatedalullinitsmulti-modalityupdate.However,webelievethemulti-modalitymarket,andespeciallythevideomodelmarket,bearssubstantialcommercialvaluewithlesscompetitioncomparedtoLLMmarkets.MiniMax'sH3releasedinJuly2026hasoutperformedglobalpeersandreceivedpositivefeedback.WeexpectH3tore-accelerate

MiniMax'smulti-modalityARRgrowth.

4.Positivefinancingcycles:OnJuly10,MiniMaxannouncedplanstoconductanequityplacementbyraisingHK$9.5bn(US$1.2bn)with10.19%dilution.Atthesametime,itlauncheda1-yearconvertiblebondissuancetoraiseHK$6.5