Ling-3.0-flash-VL Uyasebenza

I-Ling-3.0-flash-VL iyimodeli yokuqala enkulu yomdabu ye-multimodal evulekile ochungechungeni lwe-Bailing lwe-Ant InclusionAI, eyakhiwe ngokwandisa i-MoE architecture ye-Ling-3.0-flash, enamapharamitha angu-124B isiyonke futhi evula kuphela angu-5.5B ekucabangeni ngakunye. Isekela ngokwemvelo okokufaka kwezithombe nombhalo kanye namavidiyo, futhi iwindi lomongo lifinyelela ku-256K Token. I-Ling-3.0-flash-VL yenza imodeli ikwazi ukubuka, ukwenza nokuhlola ilungise ngesikhathi esisodwa njengomuntu, iguqule umsebenzi usuke ekukhiqizeni kwesikhathi esisodwa uye ekulethweni okuqhubekayo nokuthembekile.

Ukuxhumana okubukwayo ngesikhathi sangempela: vula ikhamera ukuze ubuke futhi uxoxe ngesikhathi esisodwa, ukhombe izitshalo nezilwane, izimpawu zomhlaba nezimoto, futhi uhumushe izimpawu zezilimi zakwamanye amazwe ngesikhathi sangempela.
Ukuhlela ngezimpendulo ezibukwayo: khiqiza ikhodi ye-frontend ngesithombe-skrini sekhasi lewebhu, bese uzilungisa ngokwakho ngokuyiqhathanisa nomphumela wokunikezwa, ukuze uphinde wakhe ikhasi lewebhu.
I-GUI Agent: khomba ukwakheka kwesixhumi esibonakalayo nesimo sokusebenza, uqedele imisebenzi kuwo wonke amathuluzi, njengokubheka idatha yewebhu, ukuhlela i-Excel nokukhiqiza amashadi.
Ukuhumusha imibiko yezokwelapha: khipha izinkomba ezivela ekuhlolweni kwegazi, ekuhlolweni kwe-biochemistry nasezithombeni zezokwelapha kuwo wonke amakhasi nezikhathi, ubonise ubungozi bezempilo ngezibani ezibomvu, eziphuzi neziluhlaza.
Uzakwethu wokufunda ngesikhathi sangempela: hlanganisa izithombe nezinhlaka ezibhaliwe ebhodini kumavidiyo ezifundo, uphendule imibuzo ngokushesha futhi ukhiqize izivivinyo.
Ukuhlela okuvelele kumavidiyo amade: sebenzisa inqubo ethi ”thola indawo→sika→hlola→lungisa” kumavidiyo emidlalo athatha amahora, ukhiqize ngokuzenzakalelayo iqoqo lezikhathi ezivelele.

Landela i-WeChat uphendule ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI avulekile
Isakhiwo se-MoE esisebenza kahle nesingavamile: imodeli yakhiwe ngokwandisa i-Ling-3.0-flash, enamapharamitha angu-124B isiyonke, kodwa ekucabangeni ngakunye ivula kuphela angu-5.5B. Lokhu kunciphisa kakhulu izindleko zamandla ekhompuyutha ngesikhathi kugcinwa amakhono, futhi kuyenza ifaneleke njengenodi esebenza kaningi ekuhambeni komsebenzi we-Agent.
Ukuqeqeshwa okuhlangene komdabu kwe-multimodal: izithombe, umbhalo namavidiyo kulungiswa ndawonye ekuqeqeshweni okufanayo, kunokuba kuqale kuqeqeshwe umbhalo bese kwengezwa imojuli yokubona. Ubufakazi bokusetshenziswa bubonisa ukuthi ukwengezwa kolwazi olubukwayo akwehlisi ikhono lombhalo; kunalokho kuletha ukwanda ngamaphuzu angu-4 ku-Artificial Analysis Intelligence Index.
Ukufakwa kwekhodi yokubona enokulungiswa kwanoma yisiphi isinqumo kanye ne-VideoRoPE: kusetshenziswa i-visual encoder ekwazi ukuphatha okokufaka kwezithombe nombhalo okunobukhulu obuhlukahlukene, kuhlanganiswe nekhodi yesikhundla ye-VideoRoPE ukuze kuqondwe amavidiyo omdabu, ngaphandle kokuhlukanisa ividiyo ibe amafreyimu bese kwehlisela ekulandeleni kwezithombe.
Isiqu solimi esihlanganisa i-KDA ne-Gated MLA: ingxenye yolimi igcina isakhiwo esihlanganisiwe sezendlalelo ezingu-42, kushintshaniswa indlela yokunaka ngomugqa i-KDA ne-MLA enamasango ngesilinganiso esingu-5:1. Lokhu kuhlanganisa ukusebenza kahle komongo omude nokunemba kokumodela, kusekela iwindi lomongo elingu-256K Token.
Indlela evaliwe yezimpendulo ezibukwayo: imodeli isebenzisa imodi yokwenza ephindaphindwayo ethi ”buka→yenza→hlola→lungisa”, iguqule umsebenzi usuke ekukhiqizeni kwesikhathi esisodwa uye ekuphindaphindweni kwemijikelezo eminingi: iqale iqonde ulwazi olubukwayo bese yenza, ibuke umphumela bese iwufanisa nenhloso, kuthi uma ithola ukuphambuka ilungise isu, iqinisekise ukwethembeka nokuhlolwa komphumela olethiwe.

Ukuyizama ku-inthanethi: vula iwebhusayithi esemthethweni: vakashela iwebhusayithi ye-Ling Studio ku-https://chat.ant-ling.com/chat.
Ukusetshenziswa kwamahhala: okwamanje isesikhathini sokuyizama mahhala; ngemva kokungena ngemvume ungaxoxa nayo ngokuqondile.
Ukulayisha izinto: faka umyalezo wombhalo, futhi ungalayisha izithombe noma amavidiyo ukuze uxhumane nayo nge-multimodal.
Izimo ongazizama: hlola amakhono ajwayelekile njengokuphinda wakhe ikhasi ngesithombe-skrini, ukubuza imibuzo ngevidiyo nokukhomba izinto ngesithombe.
Ukuyifaka endaweni yakini: Landa imodeli: thola izisindo ezivulekile ku-Hugging Face noma ku-ModelScope; izinguqulo ze-BF16 ne-FP8 sezitholakala, kanti inguqulo ye-FP4/INT4 iyeza maduze.
Lungisa indawo yokusebenza: lungiselela indawo yokucabanga esekela lesi silinganiso samapharamitha.
Layisha ukucabanga: sebenzisa uhlaka olujwayelekile lokucabanga ukulayisha imodeli; isekela ngokwemvelo okokufaka kwezithombe, umbhalo namavidiyo, futhi umongo ungandiswa ufinyelele ku-256K Token.
Yixhumanise nokuhamba komsebenzi: nge-API yisebenzise njengenodi esebenzayo ye-visual Agent, uyifake ekuhambeni komsebenzi wemisebenzi eminingi okuthi ”buka→yenza→hlola→lungisa”.

I-multimodal ithuthukisa umbhalo: ukuqeqeshwa okuhlangene komdabu kwe-multimodal akwehlisanga ikhono lombhalo; kunalokho inguqulo ye-multimodal inyuke ngamaphuzu angu-4 kunezinguqulo zombhalo kuphela ku-Artificial Analysis Intelligence Index.
Impendulo ebonakalayo evalekile: ngenqubo ephindaphindwayo ethi ”buka→yenza→hlola→lungisa”, umsebenzi usuka ekukhiqizeni kwesikhathi esisodwa uye ekulethweni okuqhubekayo okuthembekile nokuhlolwayo.
Ukusebenza kokucabanga okuphakeme: ngaphansi kwamapharamitha angu-124B, ukucabanga ngakunye kuvula kuphela angu-5.5B, kunciphisa kakhulu izindleko zamandla ekhompuyutha neze-Token, futhi kulungele ukubizwa kaningi ngemijikelezo eminingi.
Umongo omude kakhulu: isekela ngokwemvelo i-256K Token, ikwazi ukuphatha kalula ukuhlaziywa kwemibhalo emide, amavidiyo amade nomlando wezinyathelo eziningi ze-Agent.
Okokufaka komdabu kuzo zonke izinhlobo: kusetshenziswa i-visual encoder enokulungiswa kwanoma yisiphi isinqumo kanye ne-VideoRoPE, ukuze kusingathwe izithombe, umbhalo namavidiyo ngokwemvelo ngaphandle kokuhlukanisa amafreyimu.
Imiphumela yokuhlolwa evelele: ekuhlolweni kwe-Image-to-WebDev Arena, ithole amaphuzu angaphezu kwe-GPT-5.4 ngosayizi wamapharamitha avuliwe angu-5.5B.

Ilabhulali yamamodeli ye-HuggingFace: https://huggingface.co/inclusionAI/Ling-3.0-flash-VL
https://huggingface.co/inclusionAI/Ling-3.0-flash-VL-fp8

Ukuthuthukiswa kwe-frontend nokuphinda kwakhiwe amakhasi ewebhu: khiqiza ikhodi yekhasi lewebhu esebenzayo ngokwesithombe-skrini somklamo, bese izilungisa ngokuqhubekayo ngokuyiqhathanisa nomphumela wokunikezwa.
Ukuzenzekelayo komsebenzi we-GUI: yenza imisebenzi kuwo wonke ama-software, njengokuthatha idatha yewebhu, ukuyihlela ku-Excel nokukhiqiza amashadi ngokuzenzakalelayo.
Usizo ekuhumusheni imibiko yezokwelapha: hlanganisa imibiko yokuhlolwa kwegazi, i-biochemistry nezithombe zezokwelapha, ukhiphe izinkomba ezingavamile futhi ubonise ubungozi bezempilo ngamazinga ezibani ezibomvu, eziphuzi neziluhlaza.
Uzakwethu wokufunda emfundweni ye-inthanethi: uma ubuka ividiyo yesifundo, chaza amafomula nokubhalwe ebhodini ngokuhlanganisa izithombe, uphendule imibuzo ngokushesha futhi ukhiqize izivivinyo ezisekelwe emaphuzwini olwazi.
Ukuhlela izikhathi ezivelele emidlalweni: sebenzisa inqubo ethi ”thola indawo→sika→hlola→lungisa” kumavidiyo emidlalo amade athatha amahora, ukhiqize ngokuzenzakalelayo iqoqo lezikhathi ezinhle nezihamba kahle.