I-Spark X2.5 Uyasebenza

I-Spark X2.5 iwuchungechunge lwamamodeli amakhulu ajwayelekile aphambili akhiqizwe yi-iFlytek, oluhlanganisa imodeli eyisisekelo ye-MoE esebenza efwini (293B/30B) kanye namamodeli omthombo ovulekile asebenza emishinini, i-Spark X2.5-4B ne-Spark X2.5-1.7B. Amamodeli aqeqeshwe ngokuphelele kusetshenziswa amandla ekhompyutha akhiqizwe eChina, okwenza lonke uhlelo, kusukela ekuqaleni kuya ekugcineni, luzilawuleke. Izithuthukisi eziyinhloko zithinta ukukhiqizwa kwekhodi, ukubambisana kwama-agent kanye nokucabanga kwezibalo. Inguqulo esebenza emishinini isekela ngokwendabuko umongo omude kakhulu onama-Token angu-100万, futhi isebenzisa ukwakheka kokunaka okuxubile ukunciphisa izindleko zokubala. Ihambisana nezinkundla eziningi zehadiwe, okuhlanganisa i-NVIDIA ne-Huawei Ascend, futhi isekela izinhlaka ezidumile ezifana ne-vLLM ne-Ollama. Isivele itholakala nge-iFlytek Open Platform, i-Hugging Face nezinye iziteshi.

Ukuzilawula ngokuphelele eChina: Lonke uhlelo lokuqeqesha lwenziwa kusetshenziswa amandla ekhompyutha akhiqizwe eChina, kusukela kuma-chip asezingeni eliphansi kuya kumamodeli aphezulu.
Ukwakhiwa kwe-MoE okusebenza kahle: Kusetshenziswa ukwakheka kochwepheshe abaxubile okunamapharamitha angu-293B esewonke kanye namapharamitha asebenzayo angu-30B, okuhlanganisa ukusebenza okuphezulu nokusebenza kahle kokucabanga.
Ukukhiqizwa kwekhodi ngobuhlakani: Kuthuthukisa kakhulu ukuqonda nokukhiqizwa kwekhodi, futhi kudlula amamodeli anamapharamitha amaningi ekuhlolweni kwekhodi okugunyaziwe njenge-SWE-Bench.
Ukuzisebenzela kwama-Agent: Kuhlanganiswe ngokujulile nezinhlaka ezidumile zama-Agent, kusekela ukuhlela okuzimele kwezigaba eziningi, ukubiza amathuluzi nokubambisana emisebenzini eyinkimbinkimbi.
Ukucabanga kwezibalo nokunengqondo: Amakhono ezibalo nokucabanga okunengqondo athuthukiswe kakhulu, afinyelela ezingeni elihamba phambili embonini ekuhlolweni okufana ne-gaokao ne-AIME.
Umongo omude we-Token eyizigidi: Isekela ngokwendabuko umongo omude kakhulu onama-Token angu-100万. Iwukuphela kwemodeli esebenza emishinini enaleli khono, futhi ingacubungula incwadi ephelele noma ividiyo ende.
Umthombo ovulekile olula osebenza emishinini: Amamodeli asebenza emishinini e-4B ne-1.7B avulekile, asebenzisa ukwakheka kokunaka okuxubile ukuze asebenze kahle kumaselula, kumadivayisi e-IoT nakwezinye izisetshenziswa.
Ukuhambisana nayo yonke i-ecosystem yehadiwe: Isekela ngaphandle komthungo ama-chip e-NVIDIA, Huawei Ascend, Hygon, Moore Threads nezinye, futhi ihambisana nezinhlaka ezidumile ezifana ne-vLLM ne-Ollama.
Amakhono ezilimi eziningi: Isekela izilimi ezingaphezu kuka-200, ihlanganisa izilimi ezinkulu zomhlaba futhi ihlangabezane nezidingo zokuxhumana ngezilimi ezahlukene nokusatshalaliswa kwamazwe ngamazwe.

Landela i-WeChat uphendule ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI omthombo ovulekile
Ukuqeqeshwa kusetshenziswa amandla ekhompyutha akhiqizwe eChina: Lonke uhlelo lokuqeqesha, kusukela ekuqeqeshweni kwangaphambili kuya ekuqeqeshweni kwangemva kwalokho, lwenziwa kumaqoqo e-AI akhiqizwe eChina afana ne-Huawei Ascend. Lokhu kunciphisa ukuncika emandleni ekhompyutha aphesheya futhi kwenza lonke uhlelo luzilawuleke, kusukela kuma-chip kuya kumamodeli.
Ukwakhiwa kochwepheshe abaxubile be-MoE: Kusetshenziswa ukwakheka kochwepheshe abaxubile abanokusebenza okungajwayelekile. Inani lamapharamitha selilonke lingu-293B, kodwa ngesikhathi sokucabanga kuvulwa kuphela angu-30B, okugcina ukusebenza kwemodeli enkulu kuyilapho kunciphisa kakhulu izindleko zokucabanga nokusetshenziswa kwememori yevidiyo.
Indlela yokunaka exubile: Imodeli esebenza emishinini isebenzisa ukushintshana kongqimba olu-1 lokunaka okuphelele nezingqimba ezi-3 zokunaka ngewindi elislayidayo. Lokhu kusekela ngokwendabuko umongo omude kakhulu onama-Token angu-100万 ngezindleko eziphansi kakhulu zokubala, futhi kususa umkhawulo wokucubungula imibhalo emide emishinini.
Ukuthuthukiswa kokuqeqeshwa kwangemva kwalokho kwe-MOPD: Ngokusebenzisa ukufunda okuqiniswa ngezinga elikhulu kanye namasu afana ne-MOPD (ukudistila okuncamelayo okunezinhloso eziningi), kuthuthukiswa ngokuqondile amakhono abalulekile okukhiqiza ikhodi, ukuhlela okuzimele kwama-Agent, ukucabanga kwezibalo nokulandela imiyalelo.
Ukwakhiwa okulula okusebenzela imishini: Amamodeli e-4B ne-1.7B omthombo ovulekile athuthukiselwe ngokujulile izimo zokusetshenziswa emishinini. Azuza impendulo engaphansi kwesekhondi kumadivayisi anamandla aphansi afana nemizi ehlakaniphile namaselula, kuyilapho egcina ikhwalithi yokuqedela imisebenzi esondelene neyemodeli enkulu esebenza efwini.

Ukubiza i-API: Ngena ku-iFlytek Open Platform noma ku-Xingchen MaaS Platform ukuze uthole ukhiye, bese uxhuma ngesixhumi esibonakalayo esijwayelekile esihambisana ne-OpenAI. Okwamanje kumahhala isikhathi esilinganiselwe.
Ukulanda imodeli: Vakashela i-Hugging Face, i-ModelScope, i-GitHub nezinye iziteshi ukuze ulande amafayela esisindo se-Spark-X2.5-4B noma e-1.7B.
Ukuyifaka endaweni: Khetha indlela yokucabanga ngokuya ngehadiwe. Kuma-GPU e-NVIDIA sebenzisa i-vLLM/SGLang, kuma-NPU e-Huawei Ascend landa isithombe esisemthethweni se-Docker, kanti kuma-Apple Silicon sebenzisa i-MLX ngokuqondile.
Ukuqalisa ngokuchofoza okukodwa: Ngenisa imodeli nge-Ollama noma nge-LM Studio ukuze uqalise insizakalo yengxoxo endaweni ngaphandle kokubhala ikhodi.
Ukulungisa kahle: Layisha imodeli yomthombo ovulekile nge-LLaMA-Factory ukuze wenze ukuqeqeshwa okwengeziwe futhi uyivumelanise ngokushesha nemboni noma isimo sebhizinisi esithile.

Ukuzilawula ngokuphelele eChina: Lonke uhlelo lokuqeqesha lwenziwa kusetshenziswa amandla ekhompyutha akhiqizwe eChina afana ne-Huawei Ascend, okwenza lonke uhlelo, kusukela kuma-chip kuya kumamodeli, luzilawuleke futhi kuncishiswe ukuncika ngaphandle.
Ukwakhiwa kwe-MoE okusebenza kahle: Kusetshenziswa kuphela amapharamitha angu-30B kuma-293B esewonke, okuthola ukusebenza okusondelene namamodeli amakhulu ngezindleko eziphansi zokucabanga.
Ikhono lekhodi elihamba phambili: Idlula amamodeli asezingeni le-9B ekuhlolweni okugunyaziwe kwekhodi njenge-SWE-Bench, futhi ivelele ekusetshenzisweni kwama-algorithm nasekuqedeni ikhodi.
Ukuzisebenzela kwama-Agent: Ihambisana ngokujulile nezinhlaka ezidumile zama-Agent ezifana ne-Codex ne-OpenClaw, futhi isekela ukuhlela okuzimele kwezigaba eziningi nokubiza amathuluzi emisebenzini eyinkimbinkimbi.
Ukucabanga kwezibalo okuphambili: Imiphumela ekuhlolweni kwezibalo okufana ne-gaokao ne-AIME ifinyelela eqenjini elihamba phambili embonini, futhi amakhono okucabanga okunengqondo nawokuxazulula izinkinga athuthukiswe kakhulu.
Umongo omude we-Token eyizigidi: Isekela ngokwendabuko ukulandelana okude kakhulu okungama-Token angu-100万. Iwumkhiqizo okuwukuphela kwawo phakathi kwamamodeli asebenza emishinini onaleli khono, futhi ingacubungula incwadi ephelele noma ividiyo ende.
Umthombo ovulekile olula osebenza emishinini: Amamodeli e-4B/1.7B athuthukiselwe omaselula, amadivayisi e-IoT nezinye izisetshenziswa, afinyelela ikhwalithi yokuqedela imisebenzi esondelene neyemodeli enkulu esebenza efwini ngamapharamitha ambalwa kakhulu.
Ukuhambisana nayo yonke i-ecosystem yehadiwe: Ihambisana ngaphandle komthungo nama-chip e-NVIDIA, Huawei Ascend, Hygon, Moore Threads nezinye, futhi isekela ukusatshalaliswa ngokuchofoza okukodwa ngezinhlaka ezidumile ezifana ne-vLLM ne-Ollama.

Indawo yokugcina ye-GitHub：https://github.com/XHToken/Spark-X2.5
Iqoqo lamamodeli e-HuggingFace：https://huggingface.co/collections/XHToken/spark-x25

Ukuthuthukiswa kwezinhlelo ngobuhlakani: Ngenxa yokukhiqizwa kwekhodi okusezingeni eliphezulu nokusebenza kwayo ku-SWE-Bench, isekela lonke uhlelo lobunjiniyela besoftware, kusukela ekuhlaziyweni kwezidingo kuya ekuhlolweni nasekusetshenzisweni, futhi ingasebenza njengomhleli wezinhlelo we-AI ukuze ithuthukise kakhulu ukusebenza kahle.
Ukuzenzakalela kwama-Agent ezingeni lebhizinisi: Ihlanganiswe ngokujulile nezinhlaka ezidumile zama-Agent, isekela ukuhlela okuzimele kwezigaba eziningi, ukubiza amathuluzi nokwenza imisebenzi eyinkimbinkimbi ngomjikelezo ophelele. Ifanele ukuzenzakalela kwamahhovisi nezinqubo zebhizinisi ezinqamula amasistimu.
Ukucubungula imibhalo emide emishinini: Imodeli esebenza emishinini isekela ngokwendabuko umongo omude kakhulu onama-Token angu-100万, okwenza kube nokufundwa kwezincwadi eziphelele, ukuhlolwa kwezinkontileka ezinde nokuqonda amavidiyo amade kumadivayisi angaxhunyiwe ku-inthanethi afana nomaselula nezimoto.
Imfundo ehlakaniphile nokufundiswa kwezibalo: Njengoba ifinyelela ezingeni elihamba phambili embonini ekuhlolweni kwezibalo okufana ne-gaokao ne-AIME, ingeseka ukufundiswa komuntu ngamunye kokuxazulula izibalo, ukuqeqeshwa kokucabanga okunengqondo nokuhlelwa kwendlela yokufunda evumelana nomfundi.
Ukusatshalaliswa kwe-Xinchuang okuzilawuleka ngokuphelele: Ngokusebenzisa amandla ekhompyutha akhiqizwe eChina ukuze kuzuzwe ukuzilawula okuphelele kulo lonke uhlelo, ihlangabezana nezidingo zokusatshalaliswa kwangasese ezimbonini zikahulumeni, ezezimali, zamandla nezinye ezinezidingo eziqinile zokuphepha kwedatha nokuzimela kochungechunge lokuhlinzeka.