Muse Spark Open Uyasebenza

I-Muse Spark Open iyimodeli enkulu ye-multimodal evulekile ye-Meta, esebenzisa ukwakheka kwe-sparse mixture-of-experts, enamapharamitha angu-8160 billion kanye namapharamitha angu-42 billion asebenzayo ekuqageleni ngakunye, isekela umongo ofika ku-1 million Token futhi ingacubungula umbhalo, izithombe, amavidiyo nomsindo, igxile kakhulu ekubhaleni izinhlelo nasezintweni ezenziwa ama-Agent. Ngokusekelwe kuhlaka lwe-MLX, i-Muse Spark Open ingasebenza endaweni kuma-M5 Ultra Mac Studio amabili. Amakhono ayo ajwayelekile nawokubhala izinhlelo okwamanje asalele emuva kwamamodeli amaningana avulekile ase-China, njenge-MiMo-V2.6-Pro, GLM-5.3 kanye ne-Kimi K3.

Ukuqonda umongo omude kakhulu: Isekela iwindi lomongo elingu-1 million Token, futhi ingacubungula imibhalo emide kakhulu namakhodi amaningi ngesikhathi esisodwa.
Ukucubungula i-multimodal ngokwemvelo: Ingamukela futhi iqonde okokufaka kwezindlela ezine, okuwumbhalo, izithombe, amavidiyo nomsindo.
Ikhono lokubhala izinhlelo: Ingakwazi ukulungisa ama-Bug, ingeze imisebenzi futhi iqede ukwakhiwa kwamaphrojekthi, ikwazi nokusingatha imisebenzi eyinkimbinkimbi yobunjiniyela besoftware.
Ukwenza imisebenzi yama-Agent: Ingakwazi ukwenza imisebenzi yokucabanga eyizinyathelo eziningi isikhathi eside ngokuzimela, njengokusebenzisa i-DaVinci Resolve ukuhlela amavidiyo.
Ukusebenza endaweni kudeskithophu: Ngemva kokwenziwa ngcono ngohlaka lwe-MLX, ingasebenzisa imodeli ephelele endaweni kuma-M5 Ultra Mac Studio amabili.
Ukuhambisana nezinhlaka nama-chip amaningi: Isekela izinhlaka zokucabanga ezivamile njenge-SGLang, vLLM kanye ne-llama.cpp, futhi iyahambisana nama-chip e-NVIDIA, AMD, Apple kanye ne-Intel.
Ukusetshenziswa kwe-API efwini: Ingasebenzisa i-API ngezingxenyekazi ezifana ne-OpenRouter, Groq kanye ne-Together AI.

Landela i-WeChat uphendule ngokuthi “umthombo ovulekile” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI avulekile
Ukwakheka kwe-sparse mixture-of-experts (MoE): Imodeli inamapharamitha angu-8160 billion esewonke, kodwa ekuqageleni ngakunye kusebenza cishe amapharamitha angu-42 billion kuphela. Indlela yomzila yabela i-Token ngayinye kumanethiwekhi ochwepheshe ambalwa ukuze ayicubungule. Lokhu kugcina umthamo wemodeli ngesikhathi kunciphisa kakhulu izindleko zokubala, okwenza kube nokwenzeka ukuthi imodeli enkulu kakhulu isebenze ku-hardware elinganiselwe.
Iwindi lomongo elisezingeni lesigidi: Isebenzisa ubuchwepheshe obusebenza kahle bokucubungula ukulandelana okude, inwebe iwindi lomongo lifike ku-1 million Token. Lokhu kwenza imodeli ifake amakhodi amade kakhulu, imibhalo eminingi noma ulwazi lwamavidiyo amade ngesikhathi esisodwa, ihlangabezane nesidingo “senkumbulo” semisebenzi ethatha isikhathi eside.
Ukucabanga okusabalalisiwe ngaphansi kwesakhiwo sememori ehlanganisiwe: Isakhiwo sememori ehlanganisiwe se-Apple M5 Ultra sivumela i-GPU ehlanganisiwe ye-Mac Studio ngayinye ukuthi isebenzise imemori efinyelela ku-512GB. Amadivayisi amabili ehlangene anesamba esicishe sibe ngu-1TB, ne-memory bandwidth engu-1.2TB/s. Ngosizo lwenguqulo entsha yohlaka lwe-MLX, amadivayisi amabili angasebenza ngokubambisana njengekhompyutha enkulu eyodwa, aqukathe ngokuphelele imodeli enamapharamitha angu-8160 billion.
Ukusheshiswa ngama-kernel e-GPU enziwe ngokwezifiso: I-Meta inikeza i-SGLang ama-kernel e-Helion GPU enziwe ngokwezifiso akwazi ukuzilungiselela ngokuzenzakalela. Lokhu kufinyelela ekusheshisweni kwezinga le-kernel okuphakathi kuka-1.03 no-3.95 ezimotweni ze-NVIDIA ne-AMD, futhi kuletha ukuthuthuka kokusebenza okuphelele okucishe kube ngu-3% ku-AMD MI350.

Thola izindlela zokufinyelela: Biza imodeli ngokuqondile nge-API usebenzisa izinkundla ezifana ne-OpenRouter noma i-Together AI.
Landa izisindo zemodeli: Uma udinga ukuyifaka endaweni, vakashela i-Hugging Face ukuze ulande izisindo ezivulekile.
Khetha uhlaka lokucabanga: Khetha uhlaka olufana ne-SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp noma i-Unsloth ngokuya ngendawo ye-hardware, bese ulayisha imodeli.
Lungisa indawo ye-hardware: Efini sebenzisa ama-GPU e-NVIDIA noma e-AMD; kudeskithophu esendaweni ungasebenzisa ama-Mac Studio amabili anama-chip e-M5 Ultra ukuze usebenzise imodeli ephelele.
Bhala ikhodi yokubiza: Sebenzisa izinhlaka ezifana ne-PyTorch ukubhala iskripthi sokucabanga, ufake iziyalezo zombhalo, izithombe, amavidiyo noma umsindo, bese usetha okuqukethwe komongo okungaphakathi kuka-1 million Token.
Yenza imisebenzi yama-Agent: Sebenzisa i-API noma ukushayela okusebenza endaweni ukuze wenze ukugeleza komsebenzi kwama-Agent okuyizinyathelo eziningi, njengokubhala izinhlelo nokuhlela amavidiyo.

Isikali esihamba phambili: Inamapharamitha angu-8160 billion esewonke, futhi ekuqageleni ngakunye kusebenza amapharamitha angu-42 billion kuphela; iyimodeli enamandla kakhulu evulekile enezisindo zaseMelika eyake yaba khona kwa-Meta kuze kube manje.
Umongo omude kakhulu: Isekela iwindi lomongo elingu-1 million Token, futhi ingacubungula imibhalo emide kakhulu namakhodi amakhulu ngesikhathi esisodwa.
I-multimodal ngokwemvelo: Iphatha umbhalo, izithombe, amavidiyo nomsindo ndawonye, futhi ingashayayela ngokuqondile amathuluzi e-multimodal njengokuhlela amavidiyo.
Ingasebenza kudeskithophu endaweni: Ngemva kokwenziwa ngcono ngohlaka lwe-MLX, ingasebenza ngokuphelele kuma-M5 Ultra Mac Studio amabili, ihambisele ukugeleza komsebenzi kwama-Agent endaweni.
Imvelo evulekile: Izisindo zivuleke ngokuphelele, isekela izinhlaka eziyisikhombisa ezinkulu njenge-SGLang, vLLM kanye ne-llama.cpp, futhi iyahambisana nama-chip e-NVIDIA, AMD, Apple kanye ne-Intel.
Yenziwa ngokwezifiso kakhulu: Onjiniyela bangasebenzisa i-PyTorch ukwenza i-fine-tuning nokuthuthukiswa okuqhubekayo kwemodeli.

Iwebhusayithi yephrojekthi: https://dev.meta.ai/resources/blog/build-with-muse-spark-open
Ilabhulali yamamodeli ye-HuggingFace: https://huggingface.co/collections/meta-models/muse-spark-open

Ukuthuthukisa nokugcina i-iOS: Ilungisa ama-Bug, ingeze imisebenzi emisha futhi iqede ukwakhiwa kweklayenti elivulekile i-Ice Cubes.
Ukuhlela amavidiyo ngemuva: Isebenzisa i-DaVinci Resolve ukuhlola izinto ifreyimu ngefreyimu, ikhethe iziqephu ezisezingeni eliphezulu futhi izinqume zibe yifomethi eqondile.
Ukuhlola amakhodi amakhulu: Isebenzisa umongo we-1 million Token ukuhlaziya isakhiwo nezinkinga zendawo yokugcina amakhodi yonke ngesikhathi esisodwa.
Ukucubungula imibhalo emide kakhulu: Ifunda umbiko noma iphepha elinamakhasi angamakhulu ngesikhathi esisodwa, bese yenza isifinyezo, ukuphendula imibuzo nokuqhathanisa okuhlukahlukene.
Ukuqonda izinto ze-multimodal: Ihlaziya umbhalo, izithombe, amavidiyo nomsindo ngesikhathi esisodwa ukuze isetshenziselwe ukubuyekeza noma ukuhlaziya okuqukethwe okuphelele.