I-OpenAI ishintshe ngokungalindelekile isimo sobuhlakani bokwenziwa ngokwethulwa kwe -GPT-OSS , imodeli yayo yokuqala enkulu yomthombo ovulekile eminyakeni engaphezu kwemihlanu. Lokhu kuthuthukiswa kuphawula ukuhlukana nesu lakamuva lenkampani, elaligxile kakhulu kumamodeli azimele kanye nezinsizakalo ezivaliwe. Nge-GPT-OSS, i-OpenAI ibuyela ezimpandeni zayo, ivula umnyango wokuthi noma ubani noma inhlangano ifinyelele, isebenzise, ​​futhi yenze ngokwezifiso imodeli yolimi enamandla ngaphandle kokuthembela epulatifomu yayo yamafu noma ukuxhumana kwe-inthanethi.
Imodeli entsha itholakala ngezinguqulo ezimbili: eyodwa enamapharamitha ayizigidigidi eziyi-120.000 kanye nenye elula enama-20.000 billion . Uhlobo olunamandla kakhulu lungasebenza ku-GPU eyodwa yobungcweti enememori okungenani engu-80 GB, kuyilapho inguqulo encane iqondiswe kumakhompyutha avamile edeskithophu nama-laptop ane-RAM engu-16 GB, okuvumela ukufinyelela kumakhono e-AI athuthukile kakhulu . Zombili lezi zinguqulo zingalandwa mahhala kumapulatifomu afana ne-Hugging Face futhi zisetshenziswe kumasevisi adumile njenge-Azure noma i-AWS.
Izici zobuchwepheshe nobuchwepheshe obusetshenzisiwe
Nge -GPT-OSS , i-OpenAI yethule amamodeli amabili aklanywe ngesakhiwo esisekelwe ku-Transformer kanye ne-mixed-of-experts (MoE) . Lokhu kuthuthukisa ukusetshenziswa kwememori kanye nokubambezeleka ngokusebenzisa ingxenye ekhethiwe yochwepheshe ngethokheni yokufaka, okwandisa ukusebenza kahle ngaphandle kokudela amandla. Uhlelo lwe-120B lusebenzisa amabhlogo angu-36 futhi lusebenzisa amapharamitha ayizigidi eziyizinkulungwane ezingu-5.100 ngethokheni, kuyilapho uhlelo lwe-20B lusebenzisa amabhlogo angu-24 futhi lusebenzisa amabhiliyoni angu-3.600, okwenza lufanelekele ihadiwe engabizi kakhulu.
Womabili amamodeli aqeqeshwe kakhulu kudatha yolimi lwesiNgisi evela emikhakheni ye-STEM, izinhlelo, kanye nolwazi olujwayelekile , okuhlanganisa nesigaba esengeziwe sokulungiswa okuqondisiwe kanye nokuqinisa ukuze kuthuthukiswe ukuhambisana kwawo nemiyalelo yabantu . Amakhono awo afaka phakathi ukucabanga ngochungechunge , ikhono lokuhlukanisa izinyathelo eziphakathi ngaphambi kokunikeza impendulo, kanye nokusetshenziswa kwamathuluzi angaphandle njengokuphequlula iwebhu kanye nokusebenza kwekhodi ye-Python.
Izinketho zokutholakala, ukulayisensa nokuhlanganiswa
Enye yezinzuzo ezinkulu ze-GPT-OSS ilayisensi yayo ye-Apache 2.0 , esusa imikhawulo eminingi evamile yamanye amamodeli ayimfihlo: ivumela ukusetshenziswa kwezentengiselwano, ukusatshalaliswa kabusha, kanye nokuhlanganiswa kuzo zonke izinhlobo zamaphrojekthi, kusukela ezinkampanini ezincane kuya ezinhlanganweni ezinkulu zomphakathi noma izinhlelo ze-R&D. Izisindo zamamodeli ziyatholakala ngefomethi ye-MXFP4 futhi zifaka phakathi ukusetshenziswa kwezinkomba ku-PyTorch kanye ne-Apple Metal, kanye nokuhambisana okugcwele namathuluzi afana ne -Ollama, i-llama.cpp, i-LM Studio, kanye ne-vLLM , okwenza kube lula ukusetshenziswa kwayo kokubili ezindaweni zobungcweti nezasekhaya.
Ukufaka nokusebenzisa i-GPT-OSS-20B kulula kakhulu ngenxa yezinsiza ezifana ne-LM Studio, evumela abasebenzisi abangenalo ulwazi lokuhlela ukulanda nokuhlola imodeli endaweni besebenzisa isikhombikubona sezithombe. Ngaphezu kwalokho, i-OpenAI ishicilele izincomo kanye nemibhalo enemininingwane yokulungisa nokulungisa amamodeli, kanye nokuhlanganiswa okwenziwe ngomumo kwamapulatifomu efu kanye nawasendaweni.
Ukuphepha, ukuhlola kanye nendlela eqotho
Ukuqaliswa kwe-GPT-OSS akugxilile nje kuphela ekuvulelekeni kodwa nasekuphepheni nasekusetshenzisweni okunesibopho kwamamodeli. I-OpenAI ifake izindlela ezintsha zokubuyekeza nokuqinisekisa ukuvimbela izingozi ezihlobene nokusetshenziswa kabi, ikakhulukazi ezindaweni ezibucayi njengokuphepha kwe-inthanethi kanye nokudalwa kwama-ejenti anonya. Ngenxa yalokhu, isebenzise indlela yokuhlola ephikisanayo (iqembu elibomvu), kanye nokubuyekezwa kwangaphandle ochwepheshe abazimele kanye nezinqubo zokulingisa izimo zokuxhashazwa okukhulu.
Kanye nokukhishwa kwamamodeli, i-OpenAI iqalise inselelo yokusebenzisana kwenethiwekhi enemivuzo yezezimali kulabo ababona ubuthakathaka noma ukuziphatha okuyinkinga, befuna ukubambisana okusebenzayo emphakathini womhlaba wonke ekutholeni izingozi. Ngaphezu kwalokho, i-GPT-OSS iveza uchungechunge lwayo lokucabanga , ikhuthaza ukuhlolwa kwezimpendulo nokuqapha ukuze kuvinjelwe ukucwaswa noma amaphutha abucayi.
Ukusebenza nokuqhathanisa namanye amamodeli avulekile
Ezivivinyweni zezifundo kanye nokucabanga , i-GPT-OSS-120B isebenza ngokulinganayo, futhi idlula ngisho nakwezinye izindawo, amamodeli e-OpenAI akamuva afana ne-o4-mini, futhi isebenza kangcono kunezinye ezifana ne-DeepSeek R1 noma uchungechunge lwe-Llama ezimweni ezithile. Inguqulo ye-20B, elungiselelwe amadivayisi asekhaya, igcina ibhalansi enhle phakathi kokusebenza kahle nokusebenza , ngamaphuzu ancintisana kakhulu ezibalweni, ekubhaleni amakhodi, kanye nemisebenzi yezempilo.
I-OpenAI ixwayisa ngokuthi la mamodeli avulekile angase athambekele kakhulu "ekuboneni izinto ezingekho" (amaphutha eqiniso) uma kuqhathaniswa nalawo asekelwe efwini, uma kubhekwa umehluko ekuqeqeshweni nasekuqapheni. Kodwa-ke, abasebenzisi manje banenketho yokusebenzisa amamodeli asezingeni eliphezulu ngaphandle kokudinga izinsizakalo eziku-inthanethi.
