美国安全与新兴技术中心_开放模型在研究中的应用-2025.pdfVIP

IssueBrief

TheUseof

OpenModels

inResearch

Author

KyleMille

MiaHoffman

RebeccaGelles

October2025

ExecutiveSummary

ThereiswidespreadconsensusthatopenandfreelyavailableAImodelsbenefit

research.Yetthereisalackofempiricalevidencedetailinghowthisrelationship

manifests.Thisreportaimstofillthisgapbyinvestigatingtheuseofopenlarge

languagemodels(LLMs)inpublishedresearch,overviewingwhatorganizationsand

countriesusethemmostfrequently,andconsideringtheirwiderimpactonresearch.To

thisend,weidentifyandanalyzemorethan250publicationsthatuseopenmodelsin

waysthatrequireaccesstomodelweights,andderiveataxonomyofusecasesthat

openlyavailablemodelweightsexclusivelyorpredominantlyenable.Wethenreview

morethan130publicationsthatuseclosedmodelstocompareusecaseswhenmodel

weightsareandarenotopenlyavailable.

Ouranalysisfindsthatopenmodelsenableamorediverserangeofusecasesthan

closedmodels.Oftheeighthigh-levelusecasesforAImodelsweidentified,fiveare

exclusivelyenabledbyaccesstomodelweights,twopredominantlyrequireweights,

andonedoesnotrequireweights.Thoserequiringweightsincludecontinuously

pretrainingmodelstoexpandtheirgeneralknowledge,compressingmodelstoimprove

theirefficiency,combiningdifferentmodelsorsynchronizingtheirmodalities(e.g.,text

andimagery),andmeasuringthefunctionalityofmodelsonhardwareorthe

performanceofhardwarewhenrunningmodels.

Twousecasespredominantlyrequireaccesstoweights:fine-tuningmodelsfor

particulartasksordomains,andexaminingmodelinternalstointerprettheir

functionality.Whilesomeclosedmodelapplicationprogramminginterfaces(API)allow

fortheseusecases,theaccessofferedisgenerallyverylimitedanddoesnot,for

example,allowforcustomizedfine-tuningor