2025年OpenAI o3o4-mini技术报告英文版 .docxVIP

OpenAIo3ando4-miniSystemCard

OpenAI

April16,2025

1 Introduction

OpenAIo3andOpenAIo4-minicombinestate-of-the-artreasoningwithfulltoolcapabilities—webbrowsing,Python,imageandfileanalysis,imagegeneration,canvas,automations,filesearch,andmemory.Thesemodelsexcelatsolvingcomplexmath,coding,andscientificchallengeswhiledemonstratingstrongvisualperceptionandanalysis.Themodelsusetoolsintheirchainsofthoughttoaugmenttheircapabilities;forexample,croppingortransformingimages,searchingtheweb,orusingPythontoanalyzedataduringtheirthoughtprocess.

TheOpenAIo-seriesmodelsaretrainedwithlarge-scalereinforcementlearningonchainsofthought.Theseadvancedreasoningcapabilitiesprovidenewavenuesforimprovingthesafetyandrobustnessofourmodels.Inparticular,ourmodelscanreasonaboutoursafetypoliciesincontextwhenrespondingtopotentiallyunsafeprompts,throughdeliberativealignment[1]1.

ThisisthefirstlaunchandsystemcardtobereleasedunderVersion2ofourPreparednessFramework.OpenAI’sSafetyAdvisoryGroup(SAG)reviewedtheresultsofourPreparednessevaluationsanddeterminedthatOpenAIo3ando4-minidonotreachtheHighthresholdinanyofourthreeTrackedCategories:BiologicalandChemicalCapability,Cybersecurity,andAISelf-improvement.Wedescribetheseevaluationsbelow,andprovideanupdateonourworktomitigaterisksintheseareas.

2 ModelDataandTraining

OpenAIreasoningmodelsaretrainedtoreasonthroughreinforcementlearning.Modelsintheo-seriesfamilyaretrainedtothinkbeforetheyanswer:theycanproducealonginternalchainofthoughtbeforerespondingtotheuser.Throughtraining,thesemodelslearntorefinetheirthinkingprocess,trydifferentstrategies,andrecognizetheirmistakes.Reasoningallowsthesemodelstofollowspecificguidelinesandmodelpolicieswe’veset,helpingthemactinlinewithoursafetyex