Mellum2.1 Uyasebenza

I-Mellum2.1 iyimodeli ye-AI yokuhlela enomthombo ovulekile eyethulwe yi-JetBrains, esebenzisa ukwakhiwa kwe-mixture-of-experts okungu-12B. Ku-Mellum2.1, ukufunda okuqinisa kusuke esigabeni sokugcina sokuqeqesha kwaba ingxenye eyinhloko yokuqeqesha. Ngokuhlanganisa izigidi zokuqalisa kuma-sandbox, imodeli ingahlola ama-codebase, ihlele amafayela futhi izihlolele izinguquko, okuthuthukisa kakhulu amakhono okuhlela kwama-agent. I-throughput ye-Mellum2.1 ngaphansi komthwalo omkhulu icishe iphindwe kabili kune-Qwen3.5-9B, isekela ukuthunyelwa kwangasese futhi isiyatholakala ku-Hugging Face.

Ukuhlela kwama-agent: Ihlola ama-codebase, ihlele amafayela futhi izihlolele izinguquko; ingathola imbangela yokwehluleka kwezivivinyo, ibhale iziphakamiso zokulungisa futhi iziqinisekise.
Ukuqedela nokukhiqiza ikhodi: Iqhubeka nokugxila kwe-Mellum2, inikeze ukukhiqizwa kwekhodi okusheshayo namakhono okuphendula imibuzo ezinqubweni zokuhlela.
Ukusebenzisa amathuluzi: Ingasebenzisa amathuluzi angaphandle ukuqedela imisebenzi, ilungele ukwakha amasistimu ama-agent e-AI.
Ukucabanga okuvamile: Ingaphendula imibuzo yansuku zonke futhi ixazulule izinkinga zezibalo nengqondo ngesinyathelo ngesinyathelo.
Ukuthunyelwa kwangasese: Isekela ukusebenza endaweni noma engqalasizinda ephethwe yinhlangano, ikhodi nedatha kugcinwe ngokuphelele ngaphakathi kwebhizinisi.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lezingxoxo lamaphrojekthi e-AI anomthombo ovulekile
Ukwakhiwa kwe-mixture-of-experts (MoE): I-Mellum2.1 isebenzisa ukwakhiwa kwe-mixture-of-experts okunamapharamitha angu-12B, kodwa ekucatshangweni ngakunye kuvulwa kuphela amapharamitha angu-2.5B. Imodeli yakhiwe ngamanethiwekhi amancane amaningi “ongoti”; okokufaka kudluliselwa kwabanye balaba ngoti kuphela ngokusebenzisa indlela yokuqondisa, ngaleyo ndlela yehlise kakhulu inani lokubala kuyilapho igcina umthamo wemodeli, okwenza kube nokucatshangwa okunezindleko eziphansi kanye ne-throughput ephezulu.
Ukufunda okuqinisa ngezinga elikhulu (RL): Ukuthuthukiswa okuyinhloko kwale nguqulo kusekuhambeni kokuqeqesha: i-RL isuke esigabeni esifushane sokugcina ngemva kokuqeqeshwa kwangaphambilini yaba ingxenye eyinhloko yokuqeqesha. Ithimba lengeze imisebenzi yezibalo, yokuhlela emincintiswaneni, yesayensi, yokusebenzisa amathuluzi neyobunjiniyela besoftware. Ingqalasizinda ezakhelwe ngaphakathi yethula ama-sandbox ayizigidi ngesikhathi sokuqeqesha futhi ihlanganisa izinkulungwane zezindawo zangempela, ukuze imodeli ifunde ngamaphutha nangamasignali emivuzo amakhono okuhlola, ukuhlela nokuqinisekisa adingwa yi-agent ngokuxhumana nezimo zangempela zokuhlela.
Ukuhlunga idatha ngokuqinile: Amasethi edatha avulekile asetshenziswa ekuqeqesheni ahlungwa ngokuphelele ngaphambi kokufakwa, kususwe imisebenzi enezinkinga zezivivinyo, enezimpendulo ezingenakuqinisekiswa noma enobunzima obungafanele. Lokhu kuqinisekisa ukuthi amasignali okuqeqesha e-RL athembekile futhi kuvimbela imodeli ekufundeni empendulweni engalungile.
Ukubikezela ama-Token amaningi (MTP): I-MTP yenza imodeli ibikezele ama-Token amaningi alandelayo ekuhambeni okukodwa kwe-forward. Esimweni sesicelo esisodwa, ijubane likhuphuka cishe ngo-1.6, futhi lihambisana nokwakhiwa kwe-MoE okuvula amapharamitha ambalwa, kwenza i-throughput ye-Mellum2.1 ngaphansi komthwalo omkhulu icishe iphindwe kabili kune-Qwen3.5-9B.

Vakashela ikhasi lemodeli: Vula ikhasi lenhlangano esemthethweni ye-JetBrains ku-Hugging Face bese uthola indawo yokugcina yemodeli ye-Mellum2.1.
Landa izisindo zemodeli: Ekhasini lekhadi lemodeli, chofoza ukulanda noma udonse amafayela emodeli ngokwemigomo yelayisensi (uma kudingeka, qala uvume isivumelwano sokusebenzisa).
Khetha uhlaka lokuthunyelwa: Sebenzisa i-Transformers noma i-vLLM ekuthumeleni; eyokuqala ilungele ukuqinisekisa okulula, eyesibili ilungele izinsizakalo zokukhiqiza ezinokuxhumana okuningi ngesikhathi esisodwa.
Lungisa indawo yokucatshangwa: Lungisa indawo ye-GPU (izilinganiso ezisemthethweni zisekelwe ku-H200 eyodwa), layisha imodeli bese usetha umongo onamapharamitha angu-131K.
Vula izinketho zokusheshisa: Landela imiyalelo esemthethweni yokulungisa ukubikezela ama-Token amaningi (MTP) ukuze ukhuphule ijubane lesicelo esisodwa cishe ngo-1.6.
Thumela isicelo sokucatshangwa: Sebenzisa i-API ukuthumela iziqondiso zokukhiqiza ikhodi, ukuhlela ezingeni le-repository noma imisebenzi ye-agent, bese uthola okukhiqizwe yimodeli.

Amakhono avelele okuhlela kwama-agent: Ngemva kokuqeqeshwa okukhulu kwe-RL ezindaweni zangempela, ingahlola ama-codebase ngokuzimela, ihlele amafayela futhi iqinisekise izinguquko, ithole imbangela yokwehluleka kwezivivinyo futhi ibhale iziphakamiso zokulungisa.
Ijubane lokucatshangwa elihamba phambili embonini: I-throughput ngaphansi komthwalo omkhulu icishe iphindwe kabili kune-Qwen3.5-9B, okuyenza ibe phakathi kwamamodeli ashesha kakhulu ezingeni layo.
Impendulo esheshayo ngesicelo esisodwa: Ubuchwepheshe bokubikezela ama-Token amaningi (MTP) bukhuphula ijubane lesicelo esisodwa cishe ngo-1.6.
Ukwakhiwa okusebenza kahle: Ukwakhiwa kwe-12B MoE kuvula kuphela amapharamitha angu-2.5B, kufinyelele ukusebenza okuphezulu ngezindleko eziphansi zamandla okubala.
Ikhwalithi ephezulu yedatha yokuqeqesha: Yonke idatha yokuqeqesha ihlungwa ngokuqinile, kususwe imisebenzi enezinkinga ukuze kuqinisekiswe amasignali okufunda athembekile.
Amakhono aphelele aqinile: Ithuthuka kuzo zonke izindawo, okuhlanganisa ukuhlela, ukuhlela emincintiswaneni, izibalo, ukusebenzisa amathuluzi nolwazi olujwayelekile.
Ilungele ukuthunyelwa kwangasese: Isekela ukusebenza endaweni noma engqalasizinda ephethwe yinhlangano, ikhodi nedatha kulawulwa ngokuphelele yibhizinisi, ilungele izimo zebhizinisi ezibucayi.

Iwebhusayithi yephrojekthi: https://blog.jetbrains.com/ai/2026/10/mellum2-1-gets-to-work-a-fast-open-model-for-coding-agents/
Indawo yamamodeli ye-HuggingFace: https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking

Ukuhlela kwama-agent: Ihlola ama-codebase ngokuzimela, ihlele amafayela futhi iqinisekise izinguquko, iqedele yonke inqubo kusukela ekutholeni ukwehluleka kwezivivinyo kuya ekulungiseni.
Umsizi wangasese webhizinisi: Ufakwa kuseva yebhizinisi endaweni, uqinisekise ukuthi ikhodi nedatha akuphumi kunethiwekhi yangaphakathi.
Ukuhlanganiswa okujulile ne-IDE: Ufakwa kuma-IDE e-JetBrains noma kumathuluzi ezinkampani zangaphandle, unikeze ukuqedela ikhodi nokuphendula okuhlakaniphile.
Ukuqondisa ama-sub-agent: Ngenxa ye-latency ephansi ne-throughput ephezulu, wenza imisebenzi engaphansi noma aqondise izicelo ezinhlelweni ezinama-agent amaningi.
Ukucabanga kwezibalo nengqondo: Ucubungula ukuxazulula izibalo, imincintiswano yama-algorithm nemibuzo yansuku zonke, abe namakhono omsizi ojwayelekile.