Imodeli lensha yekubhala ikhodi isondzela etulu ngemali lencane kakhulu, futsi inembobo yinye lebonakalako

Ngamhla ti-10 September 2026, i-Cognition — inkampani lesemuva kwe-ejenti yekubhala ikhodi i-Devin — ikhiphe imodeli lensha lebitwa ngekutsi yi-SWE-2. Sihloko ngemugca munye: itfole 50.0% ku-FrontierCode 1.1 Main uma icatsaniswa na-50.9% ye-Fable 5.1 ye-Anthropic, umehluko longaphansi kweliphuzu linye, futsi i-Cognition itsi kuyisebentisa kubita kancane nga-64% kute ufinyelele kulawo maphuzu lafanako. Kodvwa indzaba ayigcini lapho. Ku-Terminal-Bench 4, i-SWE-2 itfole kuphela 27.3%, uma icatsaniswa na-55.8% ye-Fable 5.1 kanye na-57.9% ye-GPT-6 Astra — umehluko lolinganiselwa emaphuzwini lamashumi lamatsatfu ensikeni yinye.

Lokuphikisana kuyo sifundvo sangempela kumabhizinisi, hhayi inombolo leshibhile yodvwa. Singena esigabeni lapho kucatsaniswa kwemamodeli kungasagcini ku-"nguyiphi lehlakaniphe kakhulu", kodvwa kucala ngembuto lecijile: ihlakaniphe kakhulu kuluphi luhlobo lwemsebenti, ngetiphi tindleko, futsi iphukela kuphi.

Tinombolo letishicilelwe, njengoba tinjalo

I-Cognition ishicilele litafula lekucatsanisa ngco emkhatsini we-SWE-2, imodeli yayo yesisekelo, kanye nemamodeli lamabili ezingeni laphambili. Nakhu kufundza lokuphelele:

  • FrontierCode 1.1 Main: i-SWE-2 ku-50.0%, imodeli yesisekelo i-Kimi K3 ku-44.2%, i-GPT-6 Astra ku-53.3%, i-Fable 5.1 ku-50.9%.
  • DeepSWE 1.1: i-SWE-2 ku-73.0%, ngetulu kwe-Fable 5.1 lene-67.4% futsi iseduze ne-Astra lene-74.1%.
  • Terminal-Bench 2.1: i-SWE-2 ku-92.8%, lokuphakeme kakhulu etafuleni, phambi kwa-91.4% ye-Fable 5.1 kanye na-89.9% ye-Astra.
  • Terminal-Bench 4: i-SWE-2 ku-27.3% kuphela, uma icatsaniswa na-55.8% kanye na-57.9%. Lena yinsika leyehlulwa kuyo ngalokucacile.

Etinsikeni letintsatfu kuletine, i-SWE-2 imi ngalokulinganako nemamodeli labiza kakhulu noma iphambi kwawo; ensikeni yinye iwela ewalweni. Naka futsi kutsi yehlule sisekelo sayo, i-Kimi K3, kuyo yonkhe insika — lokwentiwa ncono kuyo lokwengetwe kucecesha kwe-Cognition.

Yakhiwe njani: imodeli lenkhulu levulekile, bese kucecesha lokukhetsekile ngetulu kwayo

I-SWE-2 ayizange yakhiwe kusukela ekucaleni. Sisekelo sayo yi-Kimi K3, imodeli lenemaweights lavulekile ye-Moonshot AI lenemapharamitha langu-2.8 trillion. Loko lokwentiwa yi-Cognition kwaba kucecesha kwangemuva ngekufundza ngekucinisa lokuhloswe emsebentini munye: bunjiniyela besoftware bangempela ngekhatsi kwetindzawo tekugcina ikhodi tangempela.

Imininingwane lejabulisako yebunjiniyela kutsi imodeli iza netinga letintsatfu temizamo longatikhetsa — lisemkhatsini, liphakeme, kanye nalelisetulu kakhulu. Leti atizange ticeceshwe ngetikhatsi letintsatfu letehlukene noma njengemamodeli lamatsatfu lakhetsekile, kodvwa ngesikhatsi sinye sekucecesha kusetjentiswa inhlawulo yetindleko lefakwe ku-function yemvuzo: kuphumelela kususwe tindleko temzamo, kuphindzaphindzwe nge-coefficient lecushwa ngulizinga lemizamo. Umphumela yimodeli yinye lekwatiko kutsi kufanele yonge nini nekutsi kufanele imbe kakhulu nini, esikhundleni sekukuphoca kutsi uphatse emamodeli lamatsatfu.

Umehluko lobonakalako uyacaca: ezingeni lemizamo lelisemkhatsini, lithimba litsi imodeli yehlisa linani letinyatselo nga-58% netindleko nga-81% ngesilinganiso uma kucatsaniswa nekukhishwa kwangaphambilini kwe-SWE-1.7 ku-benchmark lefanako. Ngelulwimi lwemaphrojekthi, loku kusho kuzungeleta kancane inkinga ngaphambi kwekufinyelela esisombululweni.

Umkhawulo lobalulekile ngaphambi kwekutsi ujabule: awekho emaweights lavulekile, ayikho i-API

Loku kudzinga kushiwo ngalokucacile. I-SWE-2 akusiyo imodeli longayilandza kumaseva akho, futsi awukwati kuyibita ku-app yakho nge-API lelitimele. Isebenta ngekhatsi kweplatfomu ye-Devin kuphela: kucala ku-app yedeskithophu nakulayini wemiyalo, bese iyandziswa iye ku-Web nase-Fusion. Loku kusho kutsi temnotfo tetindleko i-Cognition letichazako tiboshelwe ekubhaliseni eplatfomini yayo, hhayi enkhululekweni yekusebentisa imodeli noma kuphi lapho ufuna khona.

Lomehluko ubalulekile kunobe ngubani locatsanisa tinketfo. Imodeli lenemaweights lavulekile ikunika bukhosi etulu kwedatha yakho nenkhululeko yekuyiphatsa ngekhatsi kweMbuso; lona ngumkhicito lovalekile lotsengwa njengensita. Nangabe sizatfu sakho sekufuna lenye indlela leshibhile bekukugcina ikhodi ngekhatsi kwesimo sakho, lesimemetelo asiyixazululi inkinga yakho. Nangabe sizatfu sakho bekukunciphisa sikweleti setinsita tekutfutfukisa, mhlawumbe singayixazulula.

Loku kusho kutsini ebhizinisini lakho ngalokusebentako

Tifundvo letintsatfu kumnikati lokhokha sikweleti sekutfutfukisa noma locabanga ngetinsita te-AI telithimba lebuchwepheshe:

  • Kukhetsekisa sekube yindlela lemfisha yekuncintisana: inkampani lencane kakhulu kunemalabhorethri laphambili itsetse imodeli levulekile, yayicecesha emsebentini munye, futsi manje iyancintisana kulowo msebenti lotsite. Lombono lofanako usebenta nakuphrojekthi yakho: imodeli lesayizi lesemkhatsini lelungiswe ngedatha yakho netinchubo takho ivamise kwehlula imodeli lenkhulu lesebenta ngaphandle kwesimo.
  • Inombolo yinye esimemetelweni ayanele nanini: nangabe ubuke kuphela i-FrontierCode, ubungaphetsa ngekutsi i-SWE-2 ingangena esikhundleni ngalokuphelele. Insika ye-Terminal-Bench 4 itsi lokuphambene nalokho kuluhlobo lolutsite lwemisebenti. Hlala ubuta: kuyiphi i-benchmark, ilinganisa kutsini, futsi umehluko ukuphi.
  • Tindleko akusilo inani lesicelo ngasinye: tindleko takho tangempela linani lemizamo kuze kube nemphumela lolungile. Imodeli leshibhile lehlala izungeleta ngaphandle kwekuphela ingabita kakhulu ngalokusebentako — futsi yile ncenye inhlawulo yetindleko ekuceceshweni kwe-SWE-2 letama kuyilungisa.

Sitiphatsa njani timemetelo letinjengalesi e-Origami

Asitiboshi tinhlelo temakhasimende kumodeli yinye. Sakha lungcweti lolusemkhatsini lolwenta kutsi imodeli ishintjwe ngemuva kwemaphazili, futsi silinganisa ngemisebenti yangempela yelikhasimende esikhundleni sema-benchmark lashicilelwe: tivumelwano tangempela, ema-invoyisi angempela, tingcoco tangempela nemakhasimende. Ngaleyo ndlela, simemetelo lesinjenge-SWE-2 siba litfuba longalihlola ngeliviki linye, hhayi sincumo lesakha kabusha luhlelo lwakho.

Sinyatselo lesisebentako nangabe unelithimba lekutfutfukisa: tsatsa imisebenti lemitsatfu yangempela endzaweni yakho yekugcina ikhodi, uyisebentise ngelithulusi lakho lamanje nangalenye indlela leshibhile, bese ulinganisa tinombolo letintsatfu kuphela — ingabe umsebenti uphumelele, kudzingeke imizamo lemingakhi, ubite malini. Ngemuva kwaloko sincumo siyacaca, ngaphandle kwekutsembela etafuleni lesimemetelo lanoma ngubani.

Imitfombo