Muse Glimmer – Imodeli yolimi enkulu enamapharamitha ayizigidi eziyizinkulungwane ezingu-30 evulwe yi-Meta Uyasebenza

IMuse Glimmer imodeli yolimi enkulu enamapharamitha ayizigidi eziyizinkulungwane ezingu-30 evulwe yi-Meta, eyenzelwe ngokukhethekile ama-Workflow e-Agent ahlala kudivayisi ubusuku nemini. Ngobuchwepheshe be-4bit quantization, imodeli ingasebenza kahle ku-GPU eyikhadi elilodwa elinememori engu-24GB noma ku-Apple Silicon, futhi ukuncipha kokusebenza ngesikhathi sokucabanga kuncane kakhulu. Ukuhlolwa kwangempela kukhombisa ukuthi imodeli ingakwazi ukuphatha izingxoxo zesikhathi sangempela nokuxhumana nama-agent ku-RTX 5090, M4/M5 Max, nakwezinye izinkundla. Ngokuhlanganisa ubuchwepheshe be-DFlash speculative decoding, ifinyelela ekusheshiseni okufika ku-3.1x, futhi ihamba phambili kunabanye abaqhudelana nabo abanosayizi ofanayo kuma-benchmark amaningi, okuhlanganisa i-MCP Atlas ne-SWE-Bench Pro.

Ukucabanga kwe-Agent endaweni: Isekela ukusebenza ngemuva ubusuku nemini, futhi ilungiselelwe ngokujulile ama-Workflow ama-agent, ngakho ingaqedela ukuhlela imisebenzi eyinkimbinkimbi ngaphandle kokuxhumeka ku-inthanethi.
Ukusetshenziswa kahle ngememori encane ye-GPU: Isebenzisa ubuchwepheshe be-4bit quantization, okwenza imodeli enamapharamitha angu-30B isebenze endaweni ku-GPU eyikhadi elilodwa elinememori engu-24GB noma ku-Mac enememori ehlanganisiwe engu-32GB.
Izingxoxo ezisebenzisanayo zesikhathi sangempela: Ku-M4 Max, M5 Max nakuma-PC asezingeni eliphezulu, inikeza izingxoxo ezibushelelezi nezimpendulo zama-Agent zesikhathi sangempela ukuze ihlangabezane nezimo ezidinga ukubambezeleka okuphansi.
I-Multimodal nokubiza amathuluzi: Isekela ama-benchmark ama-agent afana ne-MCP Atlas, futhi inamakhono okusebenzisa amathuluzi, ukukhiqiza ikhodi nokucabanga ngezinyathelo eziningi.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI avulekile
Ukwakheka nobukhulu bemodeli: IMuse Glimmer isuselwe ekwakhekeni kwe-Transformer, inamapharamitha aphelele ayizigidi eziyizinkulungwane ezingu-30. Ngokuqeqeshwa kwangemva kokwakhiwa kanye ne-instruction fine-tuning eqondiswe emisebenzini ye-Agentic, iqinisa amakhono ayo okubiza amathuluzi, ukukhiqiza ikhodi nokucabanga ngezinyathelo eziningi.
Ukucindezelwa nge-4bit quantization: Isebenzisa i-low-bit quantization ukunciphisa kakhulu ubukhulu bemodeli, okwenza imodeli enamapharamitha angu-30B isebenze endaweni ku-GPU eyikhadi elilodwa elinememori engu-24GB noma ku-Mac enememori ehlanganisiwe engu-32GB. Ukuhlolwa kukhombisa ukuthi ukuncipha kokusebenza emisebenzini yama-agent kuncane kakhulu.
I-DFlash speculative decoding: Ihlanganisa indlela ye-speculative decoding, lapho imodeli esalungiselelwe uhlaka ikhiqiza ngokushesha ama-token angaba yizinketho, bese imodeli eyinhloko iwaqinisekisa ngokuhambisana. Ku-RTX 5090 ifinyelela ekusheshiseni okufika ku-3.1x, ngesivinini sokudecoda esifinyelela ku-233 tok/s.
Uhlaka lokucabanga olusebenza ezinkundleni ezahlukene: Ku-Apple Silicon isebenzisa i-ExecuTorch, kanti kuma-NVIDIA GPU isebenzisa i-llama.cpp, ukuze kuqinisekiswe ulwazi lokucabanga lwendawo olungaguquguquki nolusebenza kahle kuma-hardware ahlukahlukene.

Ukulungisa imvelo: Qinisekisa ukuthi idivayisi yendawo ine-NVIDIA GPU enememori engaphezu kuka-24GB, noma i-Apple Silicon Mac enememori ehlanganisiwe engu-32GB (i-M4/M5 Max inconyiwe).
Ukulanda ama-weight: Vakashela i-repository ye-Hugging Face ethi meta-models/Muse-Glimmer-30B, bese ulanda amafayela emodeli ye-4bit quantization kanye namafayela okumisa.
Ukukhetha uhlaka: Abasebenzisi be-NVIDIA balayisha imodeli nge-llama.cpp; abasebenzisi be-Apple Silicon bayisebenzisa nge-ExecuTorch ukuze bathole ukusebenza okuhle kakhulu.
Ukuqalisa ukucabanga: Sebenzisa i-UI ehambisanayo yokucabanga kwasendaweni ukuze ulayishe imodeli, bese uvula i-DFlash speculative decoding ukuze uthole isivinini esiphezulu sokucabanga.
Ukuxhuma ku-Workflow ye-Agent: Ngephrothokholi ye-MCP noma nge-API yendawo, xhuma imodeli kusizindalwazi solwazi lomuntu siqu, kuma-code repository nasethuluzini, ukuze kusetshenzwe ngokuzenzakalelayo imisebenzi ehlala ngemuva.

Ukusebenza okuphezulu kakhulu endaweni: Ngemva kwe-4bit quantization, ukuncipha kokusebenza kuncane kakhulu. Emisebenzini efana ne-SWE-Bench Pro ne-DeepSearch QA, idlula kakhulu i-Gemma4-31B ne-Qwen3.6-27B.
Imfuneko ye-hardware ethengekayo: Yephula umgoqo wememori ye-GPU wokusebenzisa amamodeli amakhulu endaweni. Imemori engu-24GB ingasebenzisa imodeli enamapharamitha angu-30B, yehlise izindleko kubathuthukisi ngabanye nasemabhizinisini amancane naphakathi.
Ukugxuma okukhulu kwesivinini sokucabanga: Ngokuhlanganisa i-DFlash speculative decoding, isivinini sokudecoda ku-RTX 5090 sikhuphuka sisuka ku-74.9 tok/s siye ku-233 tok/s, okuletha ukusheshisa okungu-3.1x.
Ukuphepha nokulawuleka: Ukubala kwenzeka 100% endaweni, futhi idatha ayiphumi kudivayisi. Ekuhlolweni kokuphepha kwe-Siren AgentDojo, izinga lokuphumelela kokuhlasela lingaphansi kwalabanye abaqhudelana nabo, okuhlanganisa ukusebenza nobumfihlo.

Ilabhulali yamamodeli ye-HuggingFace: https://huggingface.co/meta-models/Muse-Glimmer-30B

Umsizi we-AI wendawo: Njenge-agent yomuntu siqu ehlala ngemuva, uphatha imisebenzi yansuku zonke efana nokuphathwa kweshejuli, ukubhala okusalungiswa kwama-imeyili nokubuyekeza ikhodi, ngaphandle kokuncika kuma-API asefu.
Izimboni ezizwela ubumfihlo: Ufanele imikhakha efana nezempilo, ezezimali nezomthetho enezidingo eziphakeme kakhulu zokuhambisana nemithetho yedatha, futhi uqinisekisa ukuthi idatha esemqoka engu-100% ihlala kudivayisi yendawo.
Uchungechunge lwamathuluzi lwabathuthukisi: Uhlanganiswa ku-IDE noma ku-terminal ukuze anikeze ukuqedelwa kwekhodi okungaxhunyiwe ku-inthanethi, ukukhiqizwa kokuhlolwa okuzenzakalelayo nokuqonda ikhodi ezingeni le-repository, futhi asekele imisebenzi yobunjiniyela besofthiwe esezingeni le-SWE-Bench.
Ukufakwa kumadivayisi asemaphethelweni: Usebenza eziteshini zokusebenza ezinememori ye-GPU engu-24GB+ noma kuma-Mac asebenza kahle kakhulu, unikeze amaqembu amancane naphakathi amandla e-AI ezinga lebhizinisi ngezindleko eziphansi.