MemT 1.0 Uyasebenza

I-MemT 1.0 iyimodeli yokuqala eyisisekelo embonini eyakhelwe inkumbulo, eyethulwe yi-Memory Tensor futhi esivulwe ngokuphelele. I-MemT 1.0 inezinguqulo ezimbili: i-Flash (amapharamitha angu-288B/asebenzayo angu-17B) ne-Nano (amapharamitha angu-36B/asebenzayo angu-3B). I-MemT 1.0 ifaka umlando wolwazi ezimweni zenkumbulo zangaphakathi ezivuselelwa ngokuqhubekayo, okwenza imodeli ifunde futhi ibhale ngqo enkumbulweni ngesikhathi sokucabanga, ngaphandle kokuncika ekusesheni kwangaphandle noma ekufakeni umbhalo ngokuphindaphindiwe. Ibe phakathi kwezindlela ezimbili ezihamba phambili zenkumbulo enamapharamitha kumabhentshimakhi enkumbulo angu-11, ijubane lokucabanga ezimweni zomlando omude lenyuka lifike cishe izikhathi ezi-4, kanti ukusetshenziswa kwesitoreji kwehla kwaba ngu-1/66, kwakha isisekelo sokwakheka kwama-Agent ahlala ehambisana nabasebenzisi futhi azithuthukise ngokuphindaphindiwe.

Isitoreji esakhelwe inkumbulo: Ifaka umlando wolwazi esimweni senkumbulo esiqhubekayo sangaphakathi kwemodeli, ukuze kusetshenziswe okuhlangenwe nakho komlando ngaphandle kokunikeza umbhalo wokuqala ngesikhathi sokubuza.
Ukuvuselelwa kwenkumbulo nge-token ngayinye: Ngemva kokuphela komjikelezo ngamunye wokuxhumana, ulwazi olusha lubhalwa ngokuzenzakalelayo enkumbulweni, ngokulawula okunembile kokusulwa, ukubhalwa nokukhohlwa.
Ukwandiswa komthamo wenkumbulo: Ngendlela exubile ye-Mixture-of-Storage, umthamo ungandiswa ngokuzimela kusetshenziswa ochwepheshe abaningi benkumbulo.
Ukuphinda kusetshenziswe ulwazi lwesikhathi eside: Ama-Agent angaphinda asebenzise izindlela zokuxazulula izinkinga, amaphethini okwehluleka nezinqubo zokuqinisekisa zangaphambilini, okwandisa izinga lempumelelo ku-SWE-bench cishe ngamaphesenti angu-3.
Ukuqhubeka kwenkumbulo phakathi kwezikhathi: Izimo zenkumbulo zingagcinwa futhi zibuyiselwe ngendlela ye-serialization, ziboshwe kumsebenzisi noma emzileni womsebenzi, ukuze kusetshenziswe isikhathi eside ngokuqhubekayo.
Ukucabanga okusebenza kahle: Ezimlandweni emide, ijubane elisuka ekuqaleni kuye ekugcineni lenyuka lifike cishe izikhathi ezi-4, kanti isitoreji esiqhubekayo sithatha u-1/66 kuphela womongo ophelele futhi asikhuli njengoba umlando ukhula.

Landela i-WeChat uphendule ngokuthi “开源” ukuze ujoyine iqembu lokuxoxa ngamaphrojekthi e-AI avulekile
Ukwakhiwa okwakhelwe inkumbulo: Amamojuli enkumbulo omdabu afakwa ngokuhambisana kumamojuli akhethiwe e-Transformer, ukuze ukunaka kwasekuqaleni nokunaka kwenkumbulo kusebenze ngokufana: okokufaka kwamanje kudlula endleleni yokuqala, kanti enye indlela ibuza isimo senkumbulo ukuze ifunde umlando. Imiphumela yalezi zindlela ezimbili iyahlanganiswa ngaphambi kokubala okulandelayo, ukuze ulwazi lokuxhumana lubambe iqhaza ngqo ekukhiqizeni i-token yamanje.
Ukwakhiwa kwenkumbulo okunezigaba ezimbili: Inkumbulo yakhiwe inkumbulo yasendaweni eguquguqukayo kanye ne-hyper-memory: eyokuqala igcina ulwazi lomlando olucindezelwe, futhi umzila ngamunye wokuxhumana ugcinwa ngokuzimela futhi uvuselelwe ngokuhamba kwesikhathi; eyesibili inikeza amapharamitha wokuqagela nawokulawula, ahlala engashintshi ngemva kokuqeqeshwa. Kokubili kusekela ukugcinwa nokusetshenziswa kolwazi.
Indlela yokuvuselela ye-Gated DeltaNet-2: I-DeltaNet-2 enamasango isetshenziswa ukuvuselela kancane kancane i-matrix yenkumbulo nge-token ngayinye ngemva kokuphela komjikelezo ngamunye wokuxhumana. Amasango amane okusula, ukubhala, ukukhohlwa nokukhipha alawula ngokuhlukana ukulungiswa kokuxhumana okudala, ukufakwa kolwazi olusha, amandla okugcina, kanye nomthelela wenkumbulo empendulweni.
Ukuqeqeshwa kwezigaba ezimbili: Ukuqeqeshwa okuqhubekayo kwangaphambili ngama-token ayizigidi eziyizinkulungwane ezingu-20 kwakha ikhono lokumodela ulimi elincike enkumbulweni. Ukuqeqeshwa kwangemva kwalokho kusetshenziswa ama-token ayizigidi eziyizinkulungwane ezingu-14, kusetshenziswa ifomethi yomongo—umbuzo—impendulo ukwenza i-KL distillation. Uthisha angafunda wonke umongo, kanti umfundi ufunda ukuziphatha kukathisha esebenzisa isimo senkumbulo kuphela, futhi kuthuthukiswa amapharamitha enkumbulo kuphela.

Thola imodeli: Landa izisindo zemodeli ze-MemT-1.0-Flash noma ze-MemT-1.0-Nano ku-GitHub noma ku-HuggingFace.
Sebenzisa imodeli: Layisha imodeli endaweni yangakini noma kuseva, bese ulungisa izinsiza zememori yevidiyo ngokwenguqulo ekhethiwe.
Bopha umzila wenkumbulo: Dala umzila ozimele wokuxhumana womsebenzisi noma womsebenzi ngamunye, bese unikeza izikhala zesimo senkumbulo ezizinikele, ukuze inkumbulo yabasebenzisi abahlukene ihlukaniswe ngokuphelele.
Bhala enkumbulweni: Ngemva komjikelezo ngamunye wengxoxo, imodeli isebenzisa umsebenzi we-commit ukuze ibhale ngokuzenzakalelayo okokufaka nempendulo yalowo mjikelezo esimweni senkumbulo esiguqukayo nge-token ngayinye.
Funda enkumbulweni: Lapho kubuzwa umbuzo emjikelezweni olandelayo, imodeli ifunda ngqo isimo senkumbulo ngesikhathi sokubala kwe-forward ukuze ikhiqize impendulo.
Ukuqhubeka phakathi kwezikhathi: Gcina isimo senkumbulo ngendlela ye-serialization kusitoreji sangaphandle. Ngokubuyisela leso simo esikhathini esilandelayo, kungaqhutshwa ulwazi oluqoqwe ngaphambilini.

Ukusebenza kwenkumbulo okuhamba phambili: Kumabhentshimakhi enkumbulo angu-11, isilinganiso samaphuzu singu-50.6 (Flash)/47.0 (Nano), okwenza ibe sendaweni yokuqala neyesibili phakathi kwezindlela zenkumbulo enamapharamitha, futhi iholele isisekelo esiqine kakhulu ngamaphuzu angu-12.5.
Ukusebenzisa umlando ngaphandle kombhalo wokuqala: Ngesikhathi sokubuza, akudingeki unikeze umbhalo womlando wokuqala; imodeli ifunda ulwazi ngqo esimweni sayo sangaphakathi senkumbulo, ngaphandle kokuncika ekusesheni kwangaphandle.
Inkumbulo ingafundwa futhi ivuselelwe ngokuqhubekayo: Ngokuqeqeshwa, imodeli ifunda ukugcina, ukuvuselela nokufunda inkumbulo. Okuhlangenwe nakho okusha kubhalwa nge-token ngayinye ngaphandle kokulungisa yonke imodeli.
Ukusebenza kahle emlandweni omude: Ngomlando wama-token angu-256K, ijubane lokucabanga elisuka ekuqaleni kuye ekugcineni lenyuka cishe izikhathi ezingu-3.99, kanti isitoreji esiqhubekayo singu-1/66 kuphela womongo ophelele futhi asikhuli njengoba umlando ukhula.
Amakhono ajwayelekile acishe angalahleki: Ngomsindo wenkumbulo ongahlobene, ukulandela imiyalelo, ukukhiqiza ikhodi nokuqonda umbhalo omude kucishe kulingane nemodeli eyinhloko yokuqala, noma kube ngcono kancane.
Ukuqhubeka kolwazi phakathi kwezikhathi: Izimo zenkumbulo zingagcinwa futhi zibuyiselwe ngendlela ye-serialization, okwenza i-Agent ihlakaniphe kakhulu futhi ishibhe kakhulu njengoba isetshenziswa.
Ubukhosi benkumbulo busehlangothini lomhlinzeki wokusetshenziswa: Umnikazi webhizinisi wemodeli uba nezimpahla zenkumbulo ezihlukile, okwenza imodeli evulekile iguquke ibe yimodeli enamandla enobuhlakani obuzimele benkampani.

Indawo yokugcina ye-GitHub: https://github.com/MemTensor/MemT-1.0
Iqoqo lamamodeli le-HuggingFace: https://huggingface.co/collections/MemTensor/memt-1
Iphepha lobuchwepheshe le-arXiv: https://arxiv.org/pdf/2610.26825

Umsizi womuntu ohambisana isikhathi eside: Ukukhumbula izintandokazi, imikhuba nezingxoxo zangaphambilini zomsebenzisi phakathi kwezikhathi, ukuze kube nokuhambisana okuqhubekayo okuqonda umsebenzisi kangcono ngokusetshenziswa ngakunye.
Ukuphinda kusetshenziswe ulwazi lwe-Programming Agent: Ukubhala izindlela zokuphinda kudalwe izinkinga, amaphethini okwehluleka nezinqubo zokuqinisekisa enkumbulweni ukuze kuqondiswe imisebenzi elandelayo.
Insizakalo yamakhasimende ehlakaniphile yebhizinisi: I-Agent yensizakalo yamakhasimende ikhumbula amarekhodi omlando okuxhumana kanye nenqubekelaphambili yesevisi yekhasimende ngalinye, ngaphandle kokubuza imibuzo efanayo futhi, kanti izimpahla zenkumbulo zigcinwa kumhlinzeki wokusetshenziswa hhayi kumkhiqizi wemodeli.
Ukusebenzisana kumaphrojekthi esikhathi eside: Ukuqongelela umongo ngokuqhubekayo kumaphrojekthi athatha izinsuku noma amasonto, ngaphandle kokuqala kabusha iphrojekthi noma ukuqala ekuqaleni lapho umngcele womongo usugcwele.
Ucwaningo lwesayensi nokuzithuthukisa okuphindaphindwayo (RSI): I-AI igcina imiphumela ewusizo yokuzama nokwehluleka ngakunye njengenkumbulo, ibe yisisekelo sokuzithuthukisa nokuthuthukisa ikhodi emjikelezweni olandelayo.