I-Spark X2.6 Uyasebenza

I-Spark X2.6 (Spark-X2.6) iyimodeli eyisisekelo ye-AI ehlanganisa amadivayisi nefu, eyethulwe yi-iFlytek futhi yaqeqeshwa kusetshenziswa amandla ekhompyutha akhiqizwe eChina ngokuphelele. Imele isu elibalulekile le-AI yaseChina kwezokuphepha nokusebenza kahle kwezindleko. I-293-A30B esefwini isekela umongo wamathokheni ayisigidi, ngokufaka okungu-0.9 yuan ngamathokheni ayisigidi, okuphansi kune-DeepSeek-V4.1-Flash; i-Spark X2.6-4B eseceleni kwedivayisi isikhishiwe njengomthombo ovulekile futhi inekhono emisebenzini yama-agent efana nemibhalo yasehhovisi nokusebenzisana namathuluzi. Uhlelo luhlukanisa ngokuzenzakalela imisebenzi phakathi kwedivayisi kuphela, ifu kuphela, noma ukubambisana kwedivayisi nefu ngokobunzima bomsebenzi nezidingo zokuphepha kwedatha, ukuze idatha ebucayi ingaphumi kudivayisi futhi imisebenzi elula yenziwe ngezindleko eziphansi. Amakhono okubhala ikhodi nawama-agent asezingeni elihamba phambili, futhi asevele esetshenziswa emikhiqizweni yasehhovisi efana ne-Loomy ne-AStudio.

Ukuqonda umongo wamathokheni ayisigidi: I-293-A30B esefwini isekela umongo wamathokheni ayisigidi, futhi ingacubungula imibhalo emide kakhulu nemisebenzi eyinkimbinkimbi enezingxoxo eziningi.
Ukukhiqiza nokulungisa ikhodi: Ibe sendaweni yokuqala ezivivinyweni zamakhodi ezifana ne-SWE Pro ne-SWE Multilingual, yafinyelela ezingeni lomhlaba wonke.
Ukwenza imisebenzi yama-agent: Isekela ukubizwa kwamathuluzi imizuliswano eminingi nokusebenzisana phakathi kwamathuluzi, kanti ama-Agent okuphathwa kwezinhlelo zamabhizinisi nokubizwa kwamathuluzi e-MCP-Atlas asezingeni elihamba phambili.
Ubuhlakani bendawo kudivayisi: I-X2.6-4B ingaqedela endaweni ngaphandle kwe-inthanethi imisebenzi yasehhovisi efana nokuhlaziya ama-imeyili, ukuhlanganisa imibiko yamasonto onke nokuhlaziya ukusebenza kwebhizinisi, ngaphandle kokuthi idatha iphume kudivayisi.
Ukubambisana kwehhovisi phakathi kwedivayisi nefu: Ihlanganisa i-Spark Loomy ne-AStudio, ehlanganisa izimo ezifana nokubhala imibiko, ukwenza ama-PPT, ukubhala ikhodi nokuhlaziya idatha.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI anomthombo ovulekile
Ukuqeqeshwa kwesikhathi eside kusetshenziswa amandla ekhompyutha akhiqizwe eChina ngokuphelele: Ngokusekelwe kunkundla ye-Flying Star No. 2 yamakhadi ayi-10,000 ekhompyutha yaseChina, kusetshenziswa ubuchwepheshe bokuqeqesha obuhambisana ngokuguquguqukayo nomsebenzi wesikhathi eside. Izinsiza zokubala zabiwa ngokuzenzakalela ngokobude bomsebenzi, kuncishiswe isikhathi sokuqeqesha sedatha engu-2000B enobude obungu-256K sisuka ezinyangeni ezingu-4.5 saya ezinyangeni ezingu-1.2, futhi kuthuthukiswe ukusebenza kahle kokuqeqesha umongo omude cishe izikhathi ezi-4. Lokhu kufezekisa ukuzimela okuphelele kusukela ekuqeqesheni kuya ekuphenduleni kuma-chip aseChina.
Ukuphendula okusabalalisiwe kwe-KV Cache enezingqimba ezine: Ukuze kubhekwane nemikhawulo yememori yevidiyo nomkhawulokudonsa kuma-chip aseChina, kwakhiwe isakhiwo sokuqala se-KV Cache enezingqimba ezine esisekelwe ku-hardware yaseChina ngokuphelele kanye nobuchwepheshe bokuphendula obuzivumelanisa nokubambisana kokugcinwa nokubala. Lokhu kukhuphula kakhulu ukusetshenziswa kabusha kwe-cache nokusebenza kokubala, kwenyuse ukusebenza kokuphendula izikhathi ezi-4.5, futhi kwenyuse izinga lokushaywa kwe-cache lisuka ku-63% laya ku-95.7%, kuxazulule inkinga yomongo omude “ongangeni nokungabaleki ngokushesha”.
Isakhiwo sokubambisana sedivayisi nefu: Kusetshenziswa indlela yokuhambisa lapho idivayisi yenza ukuhlola bese ifu libambisana nayo. Imodeli eseceleni kwedivayisi iqala ihlole ubunzima bomsebenzi nokuzwela kwedatha, bese inquma phakathi kwedivayisi kuphela, ifu kuphela, noma ukubambisana kwedivayisi nefu. Kuhlanganiswe nokuqeqeshwa kokubambisana kokuhambisa okusekelwe kumodeli yokuhlola, ukunemba kokuhambisa kukhuphuke kusuka ku-81.2% kuya ku-96.1%. Idatha iyasuswa imininingwane ebucayi ngaphambi kokuthunyelwa efwini, kuhlanganiswe ukuphepha nokusebenza kahle kwezindleko.
Ukudluliselwa kwamakhono emodeli eseceleni kwedivayisi: I-X2.6-4B ithola amakhono ngokuqeqeshwa okuhlanganisa idivayisi nefu. Ikhetha ngokuzenzakalela imodeli enguthisha futhi ilungise isisindo sokuqeqesha ngokwamakhono amanje emodeli eseceleni kwedivayisi, bese isebenzisa i-algorithm yokuqeqeshwa kwangemva kokuqeqeshwa eziphindaphindayo ezivumelana nezimo. Lokhu kwenza ukufakwa kwayo emisebenzini yama-agent eseceleni kwedivayisi kufinyelele ku-41.9%, futhi idlule amamodeli e-Qwen ne-Gemma asezingeni elifanayo emisebenzini eminingi yasehhovisi.

Chaza isimo sesidingo: Ngokobunzima bomsebenzi nokuzwela kwedatha, nquma ukuthi kusetshenziswe i-API yefu, ukufakwa kwasendaweni kudivayisi, noma ukubambisana kwakho kokubili.
Ukubiza ifu: Vakashela i-Xunfei Open Platform https://maas.xfyun.cn/modelSquare/base/2628905721051136?sort=0&userSelectCategory=%255B%255D, bese ubiza i-API yemodeli ye-293-A30B, ekhokhiswa ngo-0.9 yuan ngamathokheni ayisigidi afakiwe. Ilungele ukucabanga okuyinkimbinkimbi nemisebenzi yemibhalo emide.
Ukufakwa komthombo ovulekile kudivayisi: Landa i-X2.6-4B ku-HuggingFace (https://huggingface.co/XHToken/Spark-X2.6-4B), uyifake ku-AI PC, emotweni noma endaweni yerobhothi.
Ukusetshenziswa okuqondile ezimweni zasehhovisi: Ekusetshenzisweni kwansuku zonke, khetha i-Spark Loomy noma i-Spark AStudio, esekela ukucubungula amafayela endawo ngaphandle kwe-inthanethi.
Lungisa ukubambisana kwedivayisi nefu: Vula umsebenzi wokuhambisa phakathi kwedivayisi nefu. Uhlelo luzinqumela ngokuzenzakalela ukuthi umsebenzi uqedwe endaweni, efwini noma ngokubambisana, futhi idatha ebucayi ingakhethwa ukuthi icutshungulwe kuphela kudivayisi.
Ukufakwa kwangasese kwamabhizinisi: Amabhizinisi anezidingo eziphezulu zokuphepha kwedatha angakhetha umshini wasehhovisi we-Spark ukuze afake uhlelo ngasese endaweni.

Amandla ekhompyutha aseChina azimele futhi alawulekayo: Ngokokuqala ngqa kufezekiswe uchungechunge olugcwele kusukela ekuqeqesheni kuya ekuphenduleni kusetshenziswa ama-chip aseChina, okwenza kube imodeli enkulu ye-AI ezimele ngokuphelele kuzo zonke izigaba.
Umongo omude wamathokheni ayisigidi: I-293-A30B esefwini isekela ukufakwa kwamathokheni ayisigidi, futhi ingacubungula ngesikhathi esisodwa imibhalo emide kakhulu nemisebenzi eyinkimbinkimbi.
Ukuphathwa okuhlakaniphile kwedivayisi nefu: Ngokobunzima bomsebenzi nezidingo zokuphepha kwedatha, ikhetha ngokuzenzakalela idivayisi kuphela, ifu kuphela noma ukubambisana kwedivayisi nefu, ngokunemba kokuhambisa okufinyelela ku-96.1%.
Idatha ebucayi ayiphumi kudivayisi: Imisebenzi ebucayi ingaqedwa endaweni ngaphandle kwe-inthanethi, futhi indlela yokususa imininingwane ebucayi isiza ukuxazulula izidingo eziphuthumayo zokuhambisana kwedatha emikhakheni yezemfundo, ezempilo, ezomthetho neminye.
Amakhono ekhodi nawama-agent ahamba phambili: Ibe sendaweni yokuqala ezivivinyweni zekhodi ze-SWE Pro ne-SWE Multilingual, kanti amakhono emisebenzi yama-agent asezingeni elihamba phambili.
Amakhono aqinile futhi avulekile ohlangothini lwedivayisi: I-X2.6-4B inomthombo ovulekile futhi iyafakeka, ukufakwa kwayo emisebenzini yama-agent eseceleni kwedivayisi kufinyelela ku-41.9%, futhi idlula amamodeli e-Qwen ne-Gemma asezingeni elifanayo.

Indawo yokugcina ye-GitHub: https://github.com/XHToken/Spark-X2.6
Iqoqo lamamodeli e-HuggingFace: https://huggingface.co/collections/abenzerps/spark-x26

Izinsizakalo ze-AI zamabhizinisi ezibambisana phakathi kwedivayisi nefu: Ngokusebenzisa isakhiwo sokuhambisa phakathi kwefu nedivayisi, izicelo ezilula zicutshungulwa ngokushesha kudivayisi, kuyilapho imisebenzi eyinkimbinkimbi idluliselwa kumodeli enkulu ye-AI efwini, kwehliswe kakhulu izindleko zamandla ekhompyutha efu kuyilapho kugcinwa ulwazi oluhle lokusetshenziswa.
Ihhovisi elihlakaniphile: Sebenzisa i-Loomy ne-AStudio ukubhala imibiko, ukwenza ama-PPT, ukuhlaziya idatha nokubhala ikhodi, futhi usekele ukucubungula ama-imeyili nemibhalo yendawo ngaphandle kwe-inthanethi.
Ukuhlaziya ukusebenza kwebhizinisi: Imodeli eseceleni kwedivayisi ifunda imibiko yamasonto onke nama-imeyili endawo, ihlele ngokuzenzakalela isimo sokusebenza kwebhizinisi, ikhiqize amathebula abonakalayo, futhi imake ubungozi nemisebenzi elindile.
Ukuthuthukiswa kwekhodi: Ngokuthembela emandleni ekhodi e-SWE Pro ne-SWE Multilingual asezingeni lokuqala, isiza abathuthukisi ukukhiqiza nokulungisa ikhodi kanye nokugcina amaphrojekthi ezilimi eziningi.