Lu xew
Ekipu Qwen bu Alibaba dafa ubbeeku ofisel Qwen-Image-2.1, muy xeetu defar nataal bu am 7 milyaar ci paramet yuñ defar ngir dindi bërëb bi am ci digganté nataal yi IA defar ak defarum liggéey. Modèle bi dafay boole bind-ci-nataal defar ak soppali nataal ci benn pipeline buñ boole, dafay jàppale defar nataal bu leer, ba noppi nangu ba 10 nataali royuwaay ngir liggéey yu jafee defar. Sunu sukkandikoo ci 36Kr, xeetu model bi amna poñ bu tollu ci 60.28, di raw ay konkurent yu am xët yu tëju yu melni Nano Banana 2.0 ak GPT Image 1.5, ci noonu lañuy jëfandikoo xeetu granularite bu jaxaso ngir gëna baaxal njariñu inference.
Ekipu Qwen bu Alibaba genne na Qwen-Image-2.1 muy model bu ubbeeku, muy jéego bu am solo ci defar nataal yu xarañ yi ñu mëna jëfandikoo ci askanu defarkat yi. Modèle bi dañu ko tabax ci kaw architecture DiT bu am 20 couche ak 7 milyaari paramet ci wàllu defar ay nataal, di jàppale dayo 2K. Dayo bu ndaw bii ñu ngi ko xamee ci ni muy mëna joŋante ak model yu gëna mag te am source tëju, muy joxe point de départ bu gëna woyof ci developpeur yi soxla jëfandikoo ak personaalise jumtukaayi nataal te duñu soxla sistem yu bari yuñ moom.
Benn ci wuutale yi gëna am solo ci Qwen-Image-2.1 mooy pipeline bimu boole bind-ci-nataal ak soppali nataal. Version bii wuute na ak iteration yi njëkka am, yu daan soxla model yu wuute ngir defar ak soppali, ndax dafay may jëfandikukat yi ñu sos nataal ba noppi jëfandikoo ay soppali ci gox bi, lu melni soppali mbind, soppali expression yi, wala soppali ay mbir yuñ tànn, ci benn workflow. Modèle bi itam dafay jàppale defar nataal bu leer, di xool ci saasi ndax dafay génne nataal buñ miin wala benn bu am chaine alpha bu sukkandiko ci laaj bi, loolu mooy yombal anam wi ñuy defaree ay mbiri affiche, xëti produit, ak yeneen aplikaasioŋu jëmmal.
Modèle bi mën na jàppale ba 10 nataali royuwaay ngir dugal, loolu mooy tax ñu mëna def liggéey yu jafee xam, maanaam dañu wara boole mbir yu bari, ay araf wala ay stil. Ci misaal, jëfandikukat yi mën nañu joxe nataali yéere yu wuute, jumtukaay, ak foond ngir sos seen bu méngoo, fu mbir yépp nekk ci barab bu jaar yoon, ak stil bu jaar yoon. Ngir mëna jàppale jafe-jafe yii ci anam wu jaar yoon, Qwen-Image-2.1 dafay jëfandikoo ab jumtukaay buy bàyyi xel ci granularite bu jaxaso buy jëfandikoo mekaniism yu KV Cache ngir jëfandikoowaat xayma yi ci contexte static, wàññi xayma yu bari yu amul njariñ ak wàññi mémoire wideo bi ci diiru inference.
Sunu sukkandikoo ci 36Kr, njariñu jàngat bu ñépp bokk dafa wane ni Qwen-Nataal-2.1 moo njëkka nekk ci model yi am poñ yu tollu ci 60.28, muy ëpp Nano Banana 2.0 (59.82) ak Nataalu GPT 1.5 (59.65). Modèle bi dafay wane performance bu dëgër ci topp tegtal yi, tolluwaayu ndam ci jamono, ak njub ci soppali nataal ak produit. Jëfandikukat yi ci platform reso sosio yu melni X dañu séddoo ay resilta yu teel, di sargal ni model bi mënee jëfandikoo ay grafik yu bari mbind, ba noppi di tëye màndarga yi ci araf yi ci jamonoy soppali.
Ay leeral ci cosaan: eu.36kr.com ↗
Lu tax mu am solo
Bii génne dafa wàññi bu baax bariwaayu boole jumtukaayi nataali IA yu dëggu ci jëmmal liggéey ak liggéeyu e-commerce. Qwen-Image-2.1 dafay jàppale ci anam wu leer ay background yu leer ak soppali dëkk bi, Qwen-Image-2.1 dafay wax ci ay poñ yu metti ngir jëfandikukat yi daan soxla jéego yu bari ngir waajal IA-generated asset ngir joxe ci mujj. Naturaalu open-source bu model parametre 7B dafay may developpeur yi ñu mëna jëfandikoo ak mengale mën-mën yooyu ci gox bi wala ci infrastructure privé, wàññi dependence ci APIs yu tëju yi ak mëna wàññi njëgu liggéey bi ngir liggéey defar nataal yu bari.
Qwen-Image-2.1 biñu génne dafay wax ci jafe-jafe yi ci jëfandikoo nataali IA ngir liggéeyum jëmmal. Nataal yiñ gëna xam te IA defar dañuy faral di soxla liggéey bu bari ci loxo ngir dindi lëndëm gi, defar mbind mi, wala fexe ba mbir yu bari méngoo. Suñu boole mënin yooyu ci seen bopp, xeetu model bi dafay wàññi limu jéego yi war ngir génne mbir yu mujj, loolu mooy tax IA nekk jumtukaay bu gëna am njariñ ci defarkati grafik yi, operatëri e-commerce yi ak defarkati ëmbiit li.
Xeetu open-source bi xeetu model bi am solo la, rawatina ci mbootaay yi wara yor seeni done ak seeni infrastructure. Ak dayo parametre 7B, Qwen-Image-2.1 moo gëna mëna dem ci hardware bi ci dëkk bi wala ci environmaa cloud privé buñu ko méngale ak model yu gëna mag yu tëju. Loolu dafay may developpeur yi ñu mëna mengale model bi ci anam yu ñu ko mëna jëfandikoo, lu ci melni defar nataali produit ngir e-commerce wala defar ay jumtukaayi fësal njaay, te duñu yéem ci API yu biti yu mëna am ay àppu jëfandikoo wala jafe-jafe sutura.
Li model bi mëna jëfandikoo ba 10 nataali royuwaay ak def ay coppite yu jaar yoon, ubbi na pexe yu bees ngir nettali ay jaar-jaar yu xawa jafe ak di wane produit yi. Ci misaal, màrk yi mën nañu jëfandikoo xeetu produit bi ngir defar ci lu gaaw ay xeeti nataali produit yu am background yu wuute, ay jumtukaay, wala ay mbind yuñ ciy boole, loolu mooy gaawlu liggéeyu sos mbir yi, wàññi diir bi ak njëg yi ciy méngoo ak nataali cosaan ak liggéeyum jëmmal.
Benchmark poñ yi 36Kr xamle dañuy wane ni xeetu open-source yi leegi mën nañu méngoo wala weesu performance yi gëna mag ci xeetu closed-source yi. Coppite bi munna fitnaal ñiy joxe ay balluwaay yu tëju ngir ñu gëna baaxal seen ofrand mbaa ñu wàññi njëg yi, mu mujjee amal njariñ ekosistem IA bu gina yaatu bi ndax dina tax ñu am gis gis yu bees ak yombal jëfandikoo ci xaralay defar ay nataal.
Mekanism buy weccoo xalaat: naka lay doxee
Saytu xarala yu bees yi ci ginaaw yokkute bii ci anam wu weccoo xalaat.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
Li nga wara seetaan ci topp
Saytu ni ñuy jëfandikoo Qwen-Image-2.1 ci jumtukaayi jëmmal ak njeextalam ci njëg yi ak ci man-mani yi ci sarwiisu defar nataal yu tëju yi. Saytu benchmark yu moom seen bopp yuy saytu liñu wax ci performance, rawatina lu jëm ci njubte gi ci joxe mbind ak méngoo royuwaay yu bari, ak itam bepp coppite ci sàrti lisence model bi wala jeegoo yu askan wi di def ngir gëna jubal.
Saytu bu moom boppam ci poñ yiñ joxe ak li ñuy wax ci liggéey bi dina nekk lu am solo ngir jàngat njeextalu àdduna dëgg bu Qwen-Image-2.1. Bi 36Kr waxee ni model bi moo gëna am njariñ ci Nano Banana 2.0 ak GPT Image 1.5, njariñ yi dañu sukkandikoo ci jàngat yu ñépp am ak xalaati jëfandikukat yu njëkk yi. Yokkug test bu yeneen mbootaay ak developpeur yi di def dina tax ñu mëna wóolu model bi ak ni muy méngoo ci anam yu bari yu ñuy jëfandikoo ak ci configurations hardware yi.
Jëfandikoo Qwen-Image-2.1 ci jumtukaayi jëmmal ak platform yi dina nekk luy wane njariñam ci jëfandikoo gi. Sudee model bi mën nañu ko boole ci losisel bi siiw ci wàllu jëmmal wala jumtukaayi web, mën na nekk cër buñ miin ci jëmmal IA, di indi jafe-jafe ci ni boroom xam-xam yi di jegee defar ak soppali nataal yi. Seetal yëgle yu bawoo ci platform jëmmal yu mag yi wala projet open-source yu boole ci xeetu mbir mi.
Tontu biy joge ci ñiy defar nataal yu tëju itam dina am solo lool ñu saytu ko. Sudee Qwen-Image-2.1's performance ak disponibilite open-source dafay jur loraange yu mag ci seen posisioŋ ci marse bi, furnisër yooyu mën nañu gaawlu seen génne wala ñu yamale seen njëg ak seen man-man ngir mëna wéy di joŋante. Dinaamik bi mën na yóbbe xeetu xeetu open-source bu gëna yaatu ngir tëj bërëb bi ak ay pexe yuñ moom ci yeneen domen IA itam.
Tuning ak personalisation bu Qwen-Image-2.1 bi askan wi di dawal, dina nekk lu am solo ci yokkute. Developpeur yi mën nañu sos xeetu model buñ jagleel yenn liggéey wala anam yu ñuy jëfandikoo, lu ci melni nataali medsin, wane architecture, wala defarum modd. Coppite yooyu mën nañu yokk njariñu model bi ba noppi yóbbe ay gis-gis yu bees ci wàllu defar nataali IA.