Me ya faru
OpenAI ya buga sakamakon gwaji na farko don Jalapeño, guntun ƙwanƙwasa na al'ada ta farko. Kamfanin ya ce tsarin ya ba da 1.5 zuwa 1.9 fiye da aikin AI a kowace watt a mafi yawan kayan aiki da kuma 1.7 zuwa 3.6 sau ƙananan ƙarshen latency fiye da tsarin kwatanta a fadin GPT-OSS 120B, DeepSeek R1 670B da Kimi K2.5 1T.
OpenAI ya ce Jalapeño shine guntu na al'ada ta farko kuma sakamakon ya dogara ne akan gwada guntu tare da tsarin da aka gina kewaye da shi. Kamfanin ya kimanta tsarin akan InferenceX, ma'auni na jama'a daga SemiAnalysis wanda ke auna cikakken tsari na hidimar buƙatun AI. OpenAI ya ce ya kwatanta Jalapeño tare da tsarin AI na kasuwanci da ake samu akan wuraren aiki da suka kama daga babban abin da ake buƙata don yin amfani da ma'amala sosai, rashin amfani.
A cikin GPT-OSS 120B, DeepSeek R1 da Kimi K2.5, OpenAI ya ba da rahoton 1.5 zuwa 1.9 sau fiye da aikin AI a kowace watt a mafi girman kayan aiki da 1.7 zuwa 3.6 sau ƙananan ƙare-zuwa-ƙarshen latency fiye da tsarin kwatanta. Don nauyin aiki mai ma'amala sosai, yana ba da rahoton 2.1 zuwa 4.1 mafi girman aiki. A cikin ƙari, kamfanin ya kwatanta Jalapeño tare da GB200 don GPT-OSS 120B kuma tare da GB300 don DeepSeek R1 da Kimi K2.5. OpenAI ya ce Jalapeño ya isa iyakar Pareto a fadin wuraren aiki da aka gwada, ma'ana ya sami ingantacciyar haɗakar kayan aiki, latency da ingantaccen ƙarfi a cikin waɗannan kwatancen.
Kamfanin ya ba da rahoton cewa Jalapeño yana da ƙimar ƙarfin fakitin watt 700, yayin da aka auna ƙarfin wutar lantarki ya tsaya a ko ƙasa da watts 550 akan ayyukan da aka gwada. Kwatancen da aka buga suna amfani da ƙimar ƙarfin fakiti, gami da watts 1,200 don GB200 da watts 1,400 don GB300. Don GPT-OSS 120B, OpenAI yana ba da rahoton kusan sau 1.9 mafi girma gaɗaɗɗen gauraye alamomi a sakan daya a kowace kilowatt da sau 1.7 ƙananan latency na ƙarshe zuwa ƙarshen fiye da tsarin kwatanta. Don DeepSeek R1, yana ba da rahoton kusan sau 1.7 mafi girma mafi girman aiki a kowace watt da 3.6 ƙananan latency. Don Kimi K2.5, yana ba da rahoton kusan sau 1.5 mafi girman aiki a kowace watt da 3.4 sauƙaƙan jinkiri.
OpenAI yana ba da sakamakon ga cikakken ƙira wanda ya haɗu da guntu, ƙwaƙwalwar ajiya, cibiyar sadarwa, software da tsarin ma'auni a kusa da nauyin aikin samfurin harshe. Kamfanin ya ce gine-ginen yana kiyaye yanayin ƙirar, gami da ma'ajin KV da ake amfani da su yayin tsarawa, a cikin gida inda zai yiwu kuma yana rage jinkirin sadarwa tsakanin cores da guntu. Yana bayyana prefill azaman ƙarin ƙididdigewa da ƙididdigewa kamar yadda ƙarin ƙuntatawa ta hanyar bandwidth ƙwaƙwalwar ajiya, kuma ya ce an tsara Jalapeño don ɗaukar matakai biyu. OpenAI ya kuma ce AI ya taimaka wajen motsa guntu daga ƙirar farko zuwa tapeout a cikin watanni tara kuma ya taimaka haɓaka da'irori na lissafi, amma majiyar ba ta ba da lissafi mai zaman kansa na wancan lokacin ba ko kuma cikakken rushewar aikin ƙira mai sarrafa kansa.
Me ya sa yake da mahimmanci
Sakamakon ya nuna cewa aikin ƙaddamarwa na iya ƙara dogaro da ƙira ƙira, kwakwalwan kwamfuta, ƙwaƙwalwar ajiya, hanyar sadarwa da software azaman tsari ɗaya. Idan ribar ta riƙe a sikelin samarwa, za su iya rage ƙarfi da kayan aikin da ake buƙata don ba da amsawar AI da kuma sanya nauyin aikin wakili mai ma'amala da amsawa.
Mahimmancin mahimmanci shine yuwuwar canji a cikin yadda aka inganta kayan aikin AI. Lissafin OpenAI yana jayayya cewa babban maƙasudi na gaggawa zai iya rasa aiki lokacin da ake yin lissafi, samun damar ƙwaƙwalwar ajiya da sadarwa azaman matsaloli daban-daban. Jalapeño a maimakon haka ya haɗa waɗannan abubuwan a kusa da tsarin lokaci da tsarin motsi na bayanai na ƙirar-samfurin harshe. Wannan tsarin yana da mahimmanci saboda ba da amsa ya haɗa da sarrafa saurin gaggawa da samar da alamu a jere, yayin da tsarin wakili na iya maimaita waɗannan matakan sau da yawa a cikin ɗawainiya ɗaya.
Ribar jinkirin da aka ruwaito na iya samun tasiri mai amfani ga samfuran AI masu ma'amala idan sun tsira daga aikin samarwa. OpenAI ya ce bincike mai sauri zai iya tallafawa ƙarin wakilai masu amsawa da kuma ba da damar gudanar da ayyuka masu saurin gaske a ingantattun hanyoyin da aka haɗa a baya da sauri ko fiye da yanayin daidaitawa. Tushen baya nuna waɗannan tasirin a cikin samfurin abokin ciniki mai rai, duk da haka. Yana gabatar da sakamakon ma'auni da shirin tura abubuwan more rayuwa, don haka tasirin jama'a ya kasance mai yiwuwa.
Ƙarfin wutar lantarki kuma yana da mahimmancin aiki. OpenAI ya ce samar da ƙarin aiki mai fa'ida daga wutar lantarki iri ɗaya da kayan masarufi na iya taimaka masa hidimar buƙatu a ƙananan farashi da haɓaka ƙarfin aiki. Wannan da'awar tana da kyau a matsayin manufar kasuwanci, amma tushen baya bayar da farashin kowace tambaya, jimlar farashin tsarin, buƙatun sanyaya, kashe kuɗi, zato na amfani ko shaida cewa ma'aunin ma'auni da aka ruwaito yana fassara kai tsaye zuwa ƙananan farashin masu amfani.
Tsarin ci gaban da OpenAI ya kwatanta shi kansa sananne ne. Kamfanin ya ce samfuransa sun taimaka ƙira da haɓaka Jalapeño, yayin da ake amfani da sabbin samfuran don ingantawa da tsara shi. Using Codex with GPT-Astra, the team reportedly brought three open-weight models that were not part of the original production plan to high performance within two months. OpenAI kuma ya ce ayyukan AI da aka samar sun kasance 1.5 zuwa sau 1.8 cikin sauri fiye da aiwatar da rubuce-rubucen ɗan adam don zaɓaɓɓen kulawar GPT-OSS da gaurayawan ƙwararrun masana. Waɗannan alkalumman suna aiki ne kawai ga ɓangarorin da aka zaɓa, ba cikakkun samfura ba, kuma ba su tabbatar da cewa AI na iya ƙira da kansa ko haɓaka guntu samarwa gabaɗaya ba.
Ingantacciyar hanyar sadarwa: Yadda A zahiri yake Aiki
Bincika fasahar da ke bayan wannan ci gaban ta hanyar mu'amala.
Which component of an AI application is the machine-learning model itself?
Abin kallo na gaba
OpenAI yana shirin fara tura Jalapeño a cikin kayan aikin lissafinsa a ƙarshen 2026, amma ya ce cancantar samarwa, haɓaka software, ayyukan sikelin da ingantaccen ƙirar ƙira har yanzu suna kan ci gaba. Madogararsa baya kafa tabbaci mai zaman kansa, farashin aiki, ƙarar samarwa ko yadda guntu zai yi aiki fiye da waɗanda aka gwada.
Gwajin farko shine ko Jalapeño ya kai matakin da aka tsara na tura sojoji. OpenAI ya ce yana da niyyar fara tura guntu a cikin kayan aikin lissafin sa a ƙarshen 2026, yayin da ƙwarewar samarwa da aikin software ke ci gaba. Ƙaddamarwa zai tabbatar da cewa tsarin ya zarce zanga-zangar ma'auni, amma tushen bai ƙayyade ma'aunin farko ba, wurare, nauyin aiki ko ko masu amfani za su lura da canjin matakin samfurin.
Ana buƙatar ƙarin shaida kan sake fasalin fasalin da hanyar kwatanta. OpenAI yayi amfani da ma'auni na jama'a, amma tushen baya samar da cikakkun bayanan ma'auni, adadin kwakwalwan kwamfuta ko tsarin da aka gwada, nau'ikan software, matakan amfani, sanyaya da ikon kayan aiki, ko ingantaccen ingantaccen sakamakon kwatancen. Saboda an daidaita ma'aunin kanun labarai ta amfani da kimar ikon guntu da aka buga yayin da aka kuma ba da rahoton dorewar ikon Jalapeño daban, masu karatu yakamata su bambanta ingancin matakin kunshin daga jimlar ingancin cibiyar bayanai.
Tsarin samfurin zai zama mahimmanci. OpenAI ya ba da rahoton sakamakon GPT-OSS 120B, DeepSeek R1 670B da Kimi K2.5 1T, kuma ya ce gwajin ciki ya nuna fa'ida mai fa'ida akan samfuran OpenAI na iyaka. Madogararsa baya gano waɗannan samfuran na ciki ko buga sakamakon su. Hakanan ya ce kowane ƙarin dangin ƙirar yana buƙatar sabbin kernels da ƙayyadaddun ƙirar ƙira, wanda ke nufin sassaucin guntu na iya dogaro da ci gaba da aikin software maimakon kayan masarufi kaɗai.
OpenAI ya ce Gen 2 yana da zurfi a cikin ci gaba kuma Gen 3 yana samun tsari, yayin da kuma ya jaddada cewa za ta ci gaba da tura masu haɓakawa daga NVIDIA da sauran abokan tarayya don horarwa da ƙwarewa. Dalilin da ba a sani ba shine kawai ko Jalapeño ya fi sauri a cikin gwaje-gwajen da aka ruwaito, amma yadda ya dace da tsarin dabarun samar da ababen more rayuwa, yawan ƙarfin da zai wakilta, da kuma ko al'ummomi masu zuwa suna kiyaye ma'auni tsakanin kayan aiki, latency, inganci da tallafi don canza tsarin gine-ginen samfurin.