Muse Spark 1.3 Uyasebenza

I-Muse Spark 1.3 iyimodeli entsha yolimi olukhulu eyethulwe yi-Meta, esekelwa nguZuckerberg uqobo, futhi ibhekwa njengentuthuko enkulu ye-Meta emisebenzini yokuhlela izinhlelo nama-agent ahlakaniphile. Imodeli ivelele kakhulu ekuqondeni ikhodi nasekucubunguleni imibhalo emide, kanti ukushayelwa kwamathuluzi nokusetshenziswa kwama-Token kwehliswe kakhulu. Iphinde inikeze onjiniyela i-API ngentengo encintisanayo kakhulu.

Ukuhamba komsebenzi kwama-Agent emisebenzini emide: Isekela imisebenzi evulekile esebenza isikhathi eside, ikwazi ukuqoqa umongo ngokuzimela, ihlele imikhondo emininingwaneni engahlelekile, ithole izikhala ohlelweni bese ilulungisa ngokushesha, futhi iqhubeke nokurekhoda isimo kuze kube umsebenzi ulethwa.
Amakhono okuhlela izinhlelo athuthukisiwe: Yenzelwe imisebenzi yokuhlela izinhlelo ethatha isikhathi eside, inciphisa imijikelezo yezingxoxo engadingekile, futhi ikhiqize ikhodi ehlanzekile. Isebenza kahle kakhulu ekuqondeni ama-codebase nasekuhleleni izinhlelo nge-terminal.
Ukulandela imiyalelo eyinkimbinkimbi: Emisebenzini enezinyathelo eziningi nemibandela eminingi, iyaqhubeka nokukhumbula izidingo zangaphambilini, inciphise ukukhohlwa kwemibandela noma ukuphambuka enqubweni engxenyeni yokugcina yomsebenzi.
Ukucubungula umongo omude: Isekela iwindi lomongo lama-1M tokens futhi inikeza ukunemba okuphezulu kakhulu ekusesheni imibhalo emide.
Ukuthuthukisa ukubambisana komuntu nomshini: Uma imiyalelo ingacacile, ibuza umsebenzisi imibuzo ngokuzimisela; uma umsebenzi uvinjwa, icela usizo; futhi iqinisekisa ngaphambi kokwenza izenzo ezingenakuhlehliswa.
Ukuthuthukisa ukusebenza kahle kwezindleko: Ekuphothuleni umsebenzi ofanayo, ukushayelwa kwamathuluzi kwehla cishe ngo-20%, kanti ukusetshenziswa kwama-Token kwehla cishe ngo-25%.

Ukuqeqeshwa kwangemva nokuq扩wa ngesikhathi sokuhlolwa: Isekelwe kumodeli ye-Avocado futhi yaqeqeshwa kabusha ngemva kokuqeqeshwa kwasekuqaleni, ngesikhathi kuqiniswa nekhono lokwandisa ngesikhathi sokuhlolwa (test-time scaling). Lokhu kwenza imodeli ikwazi ukwabela izinsiza eziningi zokubala ngesikhathi sokucabanga ukuze ithuthukise ikhwalithi yemiphumela.
Ukuqeqeshwa kwama-Agent ezindaweni eziningi: Imodeli yaqeqeshwa kuma-Agent harness ahlukahlukene, ukuze iqhubeke nokusebenza ngezinga elicishe lifane noma isisuswe endaweni ethile.
Ukuqondanisa imingcele yamakhono: I-Meta yaqinisa ngokukhethekile ukuqeqeshwa kwemodeli ukuze iqonde imingcele yamakhono ayo. Lokhu kuyisiza ihlukanise ngokunembe kakhudlwana phakathi kokuthi “yazini, ayazini, ingenzani, futhi ayikwazi ukwenzani”, bese iveza izithiyo ngokucacile esikhundleni sokuqamba imiphumela.
Idatha yemisebenzi emide yokuhlela izinhlelo: Kwengezwe idatha eningi yokuqeqesha evela emisebenzini yokuhlela izinhlelo ethatha isikhathi eside. Imodeli yenzelwe kangcono izimo zobunjiniyela zansuku zonke, ukuze isebenze kahle futhi ilawuleke kakhulu ekuhambeni komsebenzi wangempela wokuthuthukisa izinhlelo.

Ukuhlela izinhlelo nge-Muse Code terminal: Qalisa i-curl -fsSL https://dev.meta.ai/install.sh | bash ku-terminal ukuze ufake i-Muse Code. Lokhu kukuvumela ukuthi ushayele i-Muse Spark 1.3 njengemodeli eyisisekelo ngqo ekuhambeni komsebenzi wokuhlela izinhlelo.
Ukuxhuma ku-Meta Model API: Onjiniyela bangayishayela ngqo nge-API esemthethweni. Intengo ejwayelekile ngu-$1.25 kuma-tokens ayisigidi okufakwayo kanye no-$4.25 kuma-tokens ayisigidi okukhishwayo.
Isaphulelo se-Contributor: Sivulwa ngokuzenzakalela ngemva kokufaka i-Muse Code, sinikeze ukusetshenziswa ngentengo ephansi engu-$0.10 kokufakwayo no-$0.20 kokukhishwayo. Ngokushintshisana nalokho, imiyalelo nekhodi ekhishwayo kungasetshenziswa yi-Meta ekuqeqesheni, kodwa umsebenzisi angaphuma ngokushintsha lesi silungiselelo.
Ukuhlanganiswa nezinkundla zezinkampani zangaphandle: Izinkundla zozakwethu ezifana ne-Vercel AI Gateway sezixhume i-Muse Spark 1.3, okwenza kube lula ukuyihlanganisa nokuyishayela ekuhambeni komsebenzi okukhona.

Amakhono okuhlela izinhlelo aphambili: Ibe ngowokuqala ezilinganisweni eziningi zokuhlela izinhlelo ezifana ne-DeepSWE, SWEAtlas ne-Terminal-Bench, idlula i-GPT-5.6 Sol ne-Claude Opus 5.
Umongo omude onamandla kakhulu: Isekela umongo wama-1M tokens, ngokunemba kokusesha imibhalo emide okungaphezu kuka-98%, idlule kakhulu izimbangi zayo.
Ukusebenza kahle kwezindleko nokusetshenziswa: Emsebenzini ofanayo, ukushayelwa kwamathuluzi kwehla cishe ngo-20% futhi ukusetshenziswa kwama-Token kwehla cishe ngo-25%. Intengo ye-API iphansi kakhulu kunamamodeli asezingeni elifanayo.
Ukuzimela kwama-Agent emisebenzini emide: Iyakwazi ukuqoqa ulwazi ngokuzimela, ithole izikhala ohlelweni bese ilulungisa, futhi igcine isimo sayo emisebenzini eyinkimbinkimbi nevulekile.
Ukubambisana komuntu nomshini okungokwemvelo: Ibuza imibuzo ngokuzimisela lapho imiyalelo ingacacile, icele usizo lapho umsebenzi uvinjwa, futhi iqinisekise ngaphambi kokwenza izenzo ezingenakuhlehliswa.
Inani elincintisana kakhulu: Intengo yohlelo olujwayelekile icishe ibe yingxenye yesihlanu yentengo ye-Claude Opus 5, kanti uhlelo lwesaphulelo se-Contributor lwehle lwaya ezingeni “elicishe lingabonakali”.

Ukuthuthukiswa kwezinhlelo okuhlakaniphile: Nge-Muse Code, ingaqedela ukubhalwa kwekhodi eyinkimbinkimbi, ilungise amaphutha kumafayela amaningi futhi iqonde ama-codebase amakhulu, kuyilapho inciphisa kakhulu ukushayelwa kwamathuluzi nokusetshenziswa kwama-Token.
Ukuhlaziywa okujulile kwemibhalo emide: Isebenzisa iwindi lomongo lama-1M tokens ukucubungula imibhalo emide enezigidi zamagama, ikhiphe ulwazi phakathi kwezahluko, yenze ukuhlaziya okuhlobene futhi iseshe ngokunembile.
Ukuphathwa kwemisebenzi yesikhathi eside ngama-Agent omuntu siqu: Ilandela imigomo yomsebenzisi ngokuqhubekayo, iqoqe ulwazi ngokuzimela, ithole izikhala ohlelweni bese ilungisa isu kuze kube umphumela wokugcina ulethwa.
Ukukhiqizwa kwemibiko yobunjiniyela eyinkimbinkimbi: Ihlanganisa idatha ehlukahlukene evela emithonjeni eminingi, njengemiphumela yokulingisa ye-CFD namamodeli e-CAD, bese ikhiqiza ngokuzenzakalela umbiko wobuchwepheshe ohlelekile ngefomethi ebekiwe.