Microsoft-Decision-1 Uyasebenza

I-Microsoft-Decision-1 iyimodeli ye-AI yesizukulwane esisha yokwenza izinqumo ngokushesha eyethulwe yi-Microsoft, eyakhelwe ngokukhethekile imisebenzi yezinqumo ehlelekile. Ikhetha isisombululo esingcono kakhulu ezinkethweni ezichazwe kusengaphambili futhi inikeze impendulo ngayinye amaphuzu wamathuba ukuze izinhlelo ezilandelayo zinqume ukuthi ziqhubeke, zizame futhi noma zidlulisele kumuntu. I-Microsoft-Decision-1 ibe sesikhundleni sokuqala ngokunemba kuzo zonke izivivinyo eziyisisekelo ezingu-36, futhi ishesha izikhathi ezingu-35 kune-GPT-6 Sol. Okokufaka kubiza u-$0.042 kuphela ngamathokheni ayisigidi, kanti okukhiphayo kumahhala. Imodeli yakhelwe phezu kwe-Qwen3.5-9B futhi isiyatholakala ku-Microsoft Foundry.

Ukukhetha isisombululo esingcono kakhulu: Ikhetha impendulo engcono kakhulu ezinkethweni ezichazwe kusengaphambili futhi ikhiphe amaphuzu wamathuba ezimpendulo ezihlukene.
Ukusabalalisa ngokomzila: Isabalalisa izicelo noma imisebenzi ngobuhlakani kumamojula noma amasevisi afanele alandelayo.
Ukuhlukanisa okuqukethwe: Ihlukanisa ngokushesha okuqukethwe okufana nombhalo; isibonelo, ithimba le-Xbox liyisebenzisa ukuhlukanisa amashumi ezinkulungwane zempendulo zemidlalo.
Ukuhlela okubalulekile kwemisebenzi: Inquma ukuhleleka kokucutshungulwa kwemisebenzi eminingi esalindile.
Ukuqinisekisa imiphumela: Ihlola ikhwalithi yemiphumela ekhiqizwe yi-AI futhi inqume ukuthi yamukelwe noma izame futhi.
Ukulawula ukuhamba komsebenzi: Njengengxenye ehlela ukuhamba komsebenzi we-Agent, inquma ukuthi kuqhubekwe, kuzanywe futhi, kuphakanyiswe, noma kudluliselwe kumuntu.

Isakhiwo esiyisisekelo: I-Microsoft-Decision-1 yakhelwe phezu kwemodeli yomthombo ovulekile i-Qwen3.5-9B. Ubukhulu bayo bamapharamitha obulinganiselayo buhlanganisa isivinini sokucabanga nezindleko zokuyifaka. I-Microsoft ihlela ukuyidlulisela esikhathini esizayo kwezinye izakhiwo eziyisisekelo, kuhlanganise nochungechunge lwayo lwe-MAI: namamodeli e-OpenAI.
Ukuqeqeshwa kwangemva kokwakhiwa okukhethekile: Imodeli yadlula ekuqeqeshweni okukhethekile kwangemva kokwakhiwa okugxile “ekulinganiseni isinqumo esisodwa”, igxile ekunikezeni amaphuzu wamathuba ezinkethweni ezichazwe kusengaphambili, ngaleyo ndlela iguqule imodeli yolimi enkulu ibe injini yezinqumo ehlelekile.
Indlela yokwenza izinqumo: Imodeli iqedela ukulinganisa nokukhetha kwesicelo ngasinye sesinqumo ngesikhathi esisodwa, inciphise kakhulu ukubambezeleka uma iqhathaniswa namamodeli olimi amakhulu asebenzisa ukucabanga okuyimizuliswano eminingi. Ngokwe-Microsoft, inkinga yokubambezeleka okunqwabelanayo lapho izinqumo ezingu-20 ezilandelanayo zengeza ama-millisecond angu-100 ngayinye incishiswa kakhulu yilesi sakhiwo.
Idizayini eqinile: Yenziwe ngokukhethekile ukuze izinqumo zizinze. Ngaphansi kwezinhlobo ezingu-8 zokuphazanyiswa kwesicelo esifanayo, izinga elijwayelekile lokushintsha kwesinqumo lingu-1.3% kuphela; lapho kuchazwa kabusha izinketho noma kuguqulwa ukuhleleka kwazo, izinga lokushintsha linguziro, okuqinisekisa ukuthi izinguquko kokokufaka azithinti ukuvumelana kwesinqumo.
Ukuhambisana nokuphepha: Imodeli yahlolwa ngokuphepha kusetshenziswa izivivinyo eziyisisekelo ezingu-11 nezicelo ezingu-5250, okuhlanganisa ukubona okuqukethwe okuyingozi, ukuhlasela kwe-jailbreak nokufakwa kwemiyalelo. Iyakwazi ukwenqaba izicelo eziyingozi ngesikhathi igcina ukusebenza okuhle emisebenzini evamile.

Ukuxhuma epulatifomu: Abathuthukisi bangayisebenzisa nge-Microsoft Foundry, ipulatifomu ye-Microsoft yokuthuthukisa i-AI.
Chaza isimo sesinqumo: Guqula umsebenzi ube yinkinga yesinqumo ehlelekile ethi “izinketho ezichazwe kusengaphambili + ukukhetha isisombululo esingcono kakhulu”, njengokusabalalisa ngomzila, ukuhlukanisa okuqukethwe nokuhlela okubalulekile.
Thumela isicelo: Thumela okuqukethwe okudinga isinqumo kumodeli ngendlela ehlelekile.
Thola amaphuzu wamathuba: Imodeli ibuyisa inketho engcono kakhulu kanye namaphuzu wamathuba ezimpendulo ezihlukene, hhayi umbhalo ovulekile.
Qalisa izenzo ezilandelayo: Sebenzisa imiphumela yamaphuzu ukunquma ukuthi kuqhubekwe, kuzanywe futhi, kuphakanyiswe noma kudluliselwe ekubuyekezweni komuntu.
Hlanganisa nokuhamba komsebenzi: Ingahlanganiswa ngaphandle kobunzima ezinhlelweni ezikhona, kuma-AI Agent nasezinhlelweni zokuhamba komsebenzi, njengengxenye eyisisekelo yokuhlela.

Isivinini esikhulu kakhulu: Ukubambezeleka kwe-P50 kushesha izikhathi ezingu-35 kune-GPT-6 Sol nezikhathi ezingu-4.5 kune-Quyet-1.0-Large, okuyenza ibe imodeli yezinqumo eshesha kakhulu ekuhlolweni kwangempela.
Ukunemba okuhamba phambili: Ibe sesikhundleni sokuqala ngokunemba kuzo zonke izivivinyo eziyisisekelo ezingu-36, ezinemibuzo ecishe ibe ngu-150000.
Izindleko eziphansi kakhulu: Okokufaka kubiza u-$0.042 kuphela ngamathokheni ayisigidi, kanti okukhiphayo kumahhala. Ezivivinyweni zangaphakathi, izindleko zemisebenzi yokuhlukanisa zehle cishe izikhathi ezingu-200 uma ziqhathaniswa ne-GPT-6 Sol.
Izinqumo ezinamaphuzu: Ikhipha amaphuzu wamathuba ezimpendulo ezihlukene, okwenza kube lula ezinhlelweni ezilandelayo ukunquma ukuthi ziqhubeke, zizame futhi noma zidlulisele kumuntu. Lokhu kulawuleka kakhulu kunokukhiqiza “njengebhokisi elimnyama”.
Ukuqina okuphezulu: Ngaphansi kwezinhlobo ezingu-8 zokuphazanyiswa, izinga elijwayelekile lokushintsha kwesinqumo lingu-1.3% kuphela; lapho izinketho zibhalwa kabusha noma zihlelwa ngokuhlukile, izinga lokushintsha linguziro.
Ukuphepha nokwethembeka: Iphumelele izivivinyo zokuphepha ezingu-11 nezicelo ezingu-5250, ikwazi ukwenqaba izicelo eziyingozi ngesikhathi igcina ukusebenza okuvamile.

Iwebhusayithi yephrojekthi: https://commandline.microsoft.com/microsoft-decision-1-model-foundry/

Ukusabalalisa ngobuhlakani ngomzila: Isabalalisa ngobuhlakani izicelo zabasebenzisi noma ithrafikhi yokuhamba kwemisebenzi kumasevisi, amamodeli noma ama-Agent afanele kakhulu alandelayo.
Ukuhlukanisa nokuqukethwe kokulinganisa: Ihlukanisa ngokushesha inqwaba yempendulo namazwana abasebenzisi futhi ibone okuqukethwe okuyingozi.
Ukuhlela okubalulekile kwemisebenzi: Ezimweni ezifana namathikithi esevisi yamakhasimende nokuphathwa kwamaphutha, inquma ukuhleleka kokucutshungulwa kwemisebenzi, ivumele ukuhlelwa okuzenzakalelayo kokuhamba komsebenzi.
Ukuqinisekisa imiphumela ye-AI: Njengomhloli wezimpendulo ze-AI, inquma ukuthi umphumela okhiqiziwe uyafaneleka yini, bese inquma ukuthi uvunyelwe, uzanywe futhi noma udluliselwe ekuphathweni okuphakeme.
Izinqumo ezingavamile nokuphakanyiswa: Ezimweni ezifana nokuphendula ezigamekweni nokulawula ikhwalithi, isebenzisa amaphuzu wamathuba ukunquma ukuthi udaba ludluliselwe yini ekubuyekezweni komuntu, kwakheke indlela ephephile yokubambisana phakathi komuntu nomshini.