Step 5 Preview Uyasebenza

I-Step 5 Preview iyimodeli eyisisekelo ephambili yesizukulwane esisha eyethulwe yi-StepFun, eyakhelwe imisebenzi ye-Agentic yomhlaba wangempela. Imodeli isebenzisa ukwakheka kwe-sparse MoE, inamapharamitha angu-600B isiyonke, kanti kuvulwa angu-27B kuphela, isekela umongo wama-Token ayisigidi futhi inamakhono emidiya eminingi ngokwemvelo. Ezivivinyweni ezihlonishwayo emhlabeni wonke, iphakathi kwamamodeli amathathu avulekile ahamba phambili, futhi ukusebenza kwayo kulandela kuphela i-GPT-6 Astra ne-Claude Opus 5; izindleko zomsebenzi owodwa zingu-1/8 kuphela zalezi zakamuva.

Ukuhlela nge-AI nobunjiniyela besoftware: Ihlanganisa imisebenzi yangempela yokulungisa amaphutha, ukuthuthukisa izici nokwenza kabusha ikhodi kumarepository ekhodi angu-553 nasezilimini zokuhlela ezingu-33, futhi ingasebenzisa amadivayisi ehadiwe angempela ukwenza ukulungisa okuqhubekayo isikhathi esingaphezu kwamahora angu-3.
Imisebenzi ezimele yesikhathi eside: Ingakwazi ukwenza imisebenzi eyinkimbinkimbi ngokuqhubekayo emahoreni angu-24, njengokuthuthukisa ama-GPU kernel kusukela ekuqaleni nokwenza ngokuzenzakalelayo izinqubo zedatha yokuqeqeshwa kwangemva kokuqeqeshwa.
Ukuthuthukiswa kwe-frontend nokudala: Ingabiza i-Blender ukuze ikhiqize amafa e-3D futhi iwaxhumanise nezinhlelo ezisebenzisanayo ze-Three.js.
Ukuhlaziywa kwezezimali: Inamakhono amathathu okusesha nokuqinisekisa ulwazi, ukwenza amamodeli okuhlola inani lezinkampani nokukhiqiza imibiko yocwaningo olunzulu, aqinisekiswe ukuhlolwa kwangaphandle kwe-FrontierFinance.
Ukuqonda umongo omude kakhulu: Isekela iwindi lama-Token ayisigidi, ikwazi ukulandelela umongo ngokuqhubekayo ochungechungeni olude lwemisebenzi futhi ibize amathuluzi kaningi ukuze iqedele umsebenzi.

Ukwakhiwa kwe-sparse MoE: Imodeli inamapharamitha angu-600B isiyonke, kodwa ekuqageleni ngakunye kuvulwa amapharamitha angu-27B kuphela. Ngokusebenzisa indlela “yokubiza ochwepheshe lapho bedingeka”, igcina amakhono asezingeni eliphezulu futhi yehlise kakhulu izindleko zokubala. Lesi yisona sizathu esikhulu esenza izindleko zomsebenzi owodwa zibe ngu-1/8 kuphela ze-Claude Opus 5.
Ukulandelela umongo omude ngokuqhubekayo: Ngewindi lomongo lama-Token ayisigidi, imodeli ingagcina ngokuphelele amarekhodi okwenziwe ngaphambilini, imiphumela yamathuluzi nempendulo yokulungisa phakathi kwemisebenzi ethatha amashumi amahora. Isebenzisa impendulo yokusebenza ukunquma isenzo esilandelayo, yenze ukuhlela kwesikhathi eside nokuphindaphinda okuqhubekayo.
Umjikelezo ophelele wokubiza amathuluzi e-Agentic: Imodeli inekhono elizimele lomjikelezo “wokuhlela—ukwenza—ukubuka—ukulungisa”. Ingafinyelela kumathuluzi angaphandle afana namachweba e-serial, amakhamera nama-API, iguqule ikhodi ngokuqhubekayo ngokusekelwe esimweni namaphutha abuyiswa indawo yangempela kuze kuqedwe umsebenzi.
Ukuhlanganiswa kwemidiya eminingi ngokwemvelo: Okokufaka kombhalo nokubonwayo kucutshungulwa ngokwemvelo kusisekelo esisodwa, okwenza imodeli iqonde ulwazi olubonakalayo olufana nemibhalo nezithombe-skrini, futhi ikwazi ukukhiqiza nokuthuthukisa okuqukethwe okubukwayo.
Indlela ye-efficiency Scaling: Iqhubeka nomqondo wobuchwepheshe ovela ku-Step 3.5 Flash. Umbuzo oyinhloko uwukuthi ukubala kungaguqulwa kanjani kube amakhono emodeli ngendlela ephumelela kakhulu. Ngokuklama ukwakheka, iqhubekisela phambili umngcele we-Pareto phakathi “kwamakhono—izindleko—ukusebenza kahle”.

Ukuyizama ku-inthanethi: Vakashela inkundla ye-Studio ku-https://studio.stepfun.com/, bhalisa bese ungena ukuze uxoxe ngqo nemodeli esipheqululini.
Ukuxhuma ku-API yaseChina: Yiya ku-https://platform.stepfun.com/ ubhalise i-akhawunti yenkundla evulekile, wakhe i-API Key, bese ubiza i-interface ye-Step 5 Preview ngokwemibhalo.
Ukuxhuma ku-API yamazwe angaphandle: Abasebenzisi abangaphandle bavakashele ku-https://platform.stepfun.ai/, babhalise ngendlela efanayo futhi bathole i-API Key ukuze bayihlanganise nezinhlelo zabo noma ukuhamba komsebenzi kwe-Agent.
Ukuxhuma emisebenzini ye-Agent: Lungiselela umongo omude namapharamitha okubiza amathuluzi ngesikhathi sokubiza i-API, ukuze imodeli yenze imisebenzi eyinkimbinkimbi ye-Agentic efana nokuhlela nokucwaninga ngezezimali.

Ibhalansi engcono kakhulu phakathi kokusebenza nezindleko: Ithole amaphuzu angu-44 ku-Global Artificial Analysis Intelligence Index, yaphuma kwabathathu abahamba phambili phakathi kwamamodeli avulekile, kanti izindleko zomsebenzi owodwa zingu-1/8 kuphela ze-Claude Opus 5, iqhubekisela phambili kakhulu umngcele we-Pareto wokusebenza kahle kobuhlakani.
Amakhono e-Agent esikhathi eside asezingeni eliphezulu: Ingakwazi ukwenza imisebenzi ezimele eqhubekayo isikhathi esingaphezu kwamahora angu-22, njengokuthuthukisa i-GPU kernel ibe ngu-508 TFLOPS, idlule i-Claude Opus 5, futhi ibonise ukuhlela okuqhubekayo namandla okuzilungisa.
Iwindi lomongo elide kakhulu: Isekela umongo wama-Token ayisigidi, ingagcina amarekhodi omlando futhi iqonde umongo ngokuqhubekayo ochungechungeni olude lwemisebenzi, ilungele izimo zokubiza amathuluzi kaningi neziyinkimbinkimbi.
Ukwakhiwa kwe-sparse MoE okusebenza kahle: Kumapharamitha angu-600B isiyonke, kuvulwa angu-27B kuphela. Lokhu kunikeza ukusebenza kwezinga eliphezulu ngezindleko eziphansi kakhulu zokuqagela, ukuze amakhono afanayo asetshenziswe emsebenzini wangempela ngezindleko eziphansi.
Ukusabalala okubanzi nokujulile kwezimo: Ibonisa ukusebenza okuvelele ekuhleleni nge-AI (amarepository angu-553, izilimi ezingu-33), ocwaningweni lwezezimali nasekuthuthukisweni kokuqukethwe kwe-frontend nokudala, inamakhono okusingatha imisebenzi ehlanganisa izimboni ezahlukene.

Iwebhusayithi esemthethweni yephrojekthi: https://www.stepfun.com/step-5-preview

Ukuhlela nge-AI nokuthuthukiswa kwesoftware: Ihlanganisa ukulungisa amaphutha, ukuthuthukisa izici, ukwenza kabusha ikhodi nokulungisa izilungiselelo zemvelo. Isekela izilimi zokuhlela ezingu-33 futhi ingasetshenziswa njengemodeli eyinhloko yansuku zonke yonjiniyela.
Imisebenzi yocwaningo nobunjiniyela yesikhathi eside: Njengokuthuthukisa i-GPU kernel ngokuqhubekayo amahora angu-24 nokwenza ngokuzenzakalelayo izinqubo zedatha yokuqeqeshwa kwangemva kokuqeqeshwa, ilungele umsebenzi wocwaningo odinga ukuphindaphinda nokusebenza okuzimele isikhathi eside.
Ukuthuthukiswa kwehadiwe nokulungisa amasistimu ashumekiwe: Ingafunda ngokuzenzakalelayo imibhalo yamadivayisi, isebenzise ama-serial port namakhamera, futhi ilungise ikhodi ngokuqhubekayo ngokusekelwe kumaphutha abuyiswa yehadiwe yangempela, njengamaphrojekthi okuguqula amabhodi okuthuthukisa e-ESP32.
Ukuhlaziywa kwezezimali nocwaningo lokutshalwa kwezimali: Kuhlanganisa ukusesha nokuqinisekisa ulwazi lwezezimali, ukwenza amamodeli okuhlola inani lezinkampani nokubhala imibiko yocwaningo olunzulu, kusekele abacwaningi bokutshalwa kwezimali nochwepheshe bezezimali.
Ukuthuthukiswa kwe-frontend nokuqukethwe kokudala: Kusukela ezindaweni zokuxhumana zamakhasi ewebhu nasekubonisweni kwedatha kuya ekukhiqizweni kwamafa e-3D nge-Blender nasezinhlelweni ezisebenzisanayo ze-Three.js, kusekela abaklami nabasebenzi bokudala ukuba basebenzise imibono yabo ngokushesha.