书生-S2 Uyasebenza

书生-S2 iyimodeli enkulu eyisisekelo ye-multimodal ekhishwe njenge-open source yi-Shanghai Artificial Intelligence Laboratory, enamapharamitha angu-397B, futhi iyimodeli enkulu eyiqonda kakhulu isayensi. Imodeli yethula amamojuli enkumbulo ochwepheshe axhunywayo ngokusebenzisa ukwakheka kwe-Memory Decoder, okwenza kube nokwandiswa okuqhubekayo kolwazi lwezizinda ngaphandle kokuqeqesha kabusha imodeli eyinhloko. Amakhono ajwayelekile ka-书生-S2 asezingeni eliphambili phakathi kwamamodeli e-open source; idlula kakhulu emisebenzini yokuhlola yesayensi efana ne-biology nezinto zokwakha, ukufaneleka kwayo ekucabangeni kwezibalo kufinyelela ezingeni le-Gemini 3.1 Pro, futhi inamakhono ocwaningo lwesayensi lwesikhathi eside kanye ne-agent ehlakaniphile.

Imibuzo nolwazi lwesayensi kanye nokucabanga: Inikeza ukuqonda nokukhiqiza okusezingeni lochwepheshe emikhakheni efana ne-biology nesayensi yezinto zokwakha, futhi idlula kakhulu amamodeli aphambili e-open source nawe-closed source.
Ukucabanga kwezibalo okuqinile kwesikhathi eside: Iyakwazi ukuqhubeka amahora amaningi icabanga futhi iqinisekisa ezilinganisweni zezibalo zokuncintisana nezibalo eziphakeme ezifana ne-IMO-Proof ne-AdvancedMathBench, ifinyelele ezingeni le-Gemini 3.1 Pro.
Ukwenza imisebenzi yocwaningo lwesayensi: Iyakwazi ukuhlaziya, ukucabanga nokwenza ngcono ngokuqhubekayo izinkinga zangempela zocwaningo, njengokuklama ama-protein binder nokwakha izakhiwo zamakristalu.
Ukwenza ngokuzenzakalelayo nge-agent ehlakaniphile: Ngomyalo owodwa ingaqedela ngokuzenzakalelayo ukusesha, ukuhlukanisa imisebenzi, ukubambisana namathuluzi nokuhlela izixazululo.
Amakhono ajwayelekile: Amakhono ajwayelekile olwazi, ikhodi nokunye asezingeni eliphambili phakathi kwamamodeli e-open source.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI open source
Ukwakhiwa kwenkumbulo ochwepheshe exhunywayo kwe-Memory Decoder: Kusetshenziswa indlela ye-Memory Decoder, ulwazi lochwepheshe lwemikhakha ehlukene lufundwa futhi lugcinwe ngamamojuli enkumbulo azimele, futhi ukuxhuma lezi zinkumbulo akudingi ukushintsha amapharamitha emodeli eyinhloko. Ngesikhathi sokucabanga, imodeli ixhumanisa ngokuguquguqukayo umnikelo wemodeli eyinhloko nowenkumbulo yoqeqesho ngokusekelwe kumongo wamanje, ukuze ulwazi lwesizinda luhlanganiswe ochungechungeni oluphelele lokucabanga.
Ukucabanga kwesikhathi eside namakhono e-agent ehlakaniphile: 书生-S2 iqinisa ikhono lokucabanga isikhathi eside ngenkathi igcina amakhono ajwayelekile, futhi isekela ukwenza imisebenzi yocwaningo esezingeni lesayensi ethatha amahora amaningi ngokusebenzisa ukuhlola, ukucabanga nokuqinisekisa okuqhubekayo. Iphinde yakha uhlaka lwe-agent ehlakaniphile olwenza imodeli iqonde izinhloso zomsebenzi, ihlukanise izinyathelo ngokuzenzakalelayo, ikhethe futhi ibambisane namathuluzi amaningi, bese ilungisa izenzo ezilandelayo ngokusekelwe empendulweni yokwenziwa komsebenzi.
Ukuthuthukiswa ngokubambisana namandla ekhompuyutha e-Ascend: Imodeli isebenzisana ngokujulile ne-ecosystem yekhompuyutha ye-Ascend, ngokuthuthukisa ngokujulile izici ezibalulekile ezifana nokubala, ukuxhumana nememori yevidiyo, futhi ihlola indlela yokuthuthukiswa ngokubambisana kwamakhono emodeli nengqalasizinda yamandla ekhompuyutha ekhiqizwa kuleli, inikeze ukwesekwa okuyisisekelo kwemisebenzi emikhulu yokubala kwesayensi.
Ukuhlolwa kokuqeqeshwa kwangaphambili endaweni efihlekile (NCP-ArchPreview): Ngokuqondiswe ekwakhiweni kwesizukulwane esilandelayo, i-Shanghai AI Laboratory iphakamise umsebenzi wokuqeqeshwa kwangaphambili endaweni efihlekile wohlobo lwe-JEPA we-Next Concept Prediction: imodeli ayisabikezeli i-Token elandelayo kuphela, kodwa isebenzisa i-Concept Module ekhethekile ukubikezela imiqondo yezincazelo ehlukene phakathi kwama-Token endaweni efihlekile, bese ibuyisela le miphumela ukuze iqondise ukukhiqizwa kwe-autoregressive okuyisisekelo.

Ukuhlola ku-inthanethi: Vakashela ku-https://chat.intern-ai.org.cn/ ukuze uxoxe ngqo futhi uzame amakhono ahlukahlukene ka-书生-S2.
Ukubiza nge-API: Bonke abasebenzisi bangasebenzisa isabelo samahhala se-API esivulekile; uma udinga isabelo esiphezulu, vakashela ku-https://internlm.intern-ai.org.cn/api/strategy ukuze uthumele isicelo.
Ukuthola imodeli: Landa izisindo zemodeli ku-HuggingFace.
Ukufakwa endaweni: Bheka i-GitHub repository ethi https://github.com/InternLM/Intern-S1 ukuze uyifake; imvelo yamandla ekhompuyutha e-Ascend isivele yathola ukulungiselelwa okujulile.

Amakhono esayensi asezingeni eliphezulu: Idlula kakhulu amamodeli aphambili akhona e-open source nawe-closed source emisebenzini yokuhlola yesayensi efana ne-biology nezinto zokwakha.
Imemori ochwepheshe exhunywayo: Ukwakhiwa kwe-Memory Decoder kusekela ukuxhuma ulwazi lwezizinda ezintsha ngaphandle kokushintsha amapharamitha emodeli eyinhloko, kunciphise kakhulu izindleko zokwandisa izizinda.
Amakhono ajwayelekile awancishiswa: Ngenkathi inweba amakhono ochwepheshe, amakhono ajwayelekile olwazi, ikhodi ne-agent ehlakaniphile ahlala esigabeni esiphambili se-open source.
Ikhono lokucabanga kwesikhathi eside: Iyakwazi ukwenza ukucabanga kwezibalo okuqinile nocwaningo amahora amaningi, ifinyelele ezingeni le-Gemini 3.1 Pro, futhi idlule kuleli zinga kwezinye izinkomba.
Ikhono lokwenza le-agent: Iyakwazi ukuhlukanisa imisebenzi eyinkimbinkimbi ngokuzenzakalelayo nokubambisana namathuluzi amaningi, futhi ngomusho owodwa ikhiqize umphumela ophelele ofana nemodeli ye-3D CAD yerokhethi i-Long March 7.

GitHub repository: https://github.com/InternLM/Intern-S1
Umtapo wamamodeli we-HuggingFace: https://huggingface.co/internlm/Intern-S2-397B

Ucwaningo nokuthuthukiswa kwesayensi yezinto eziphilayo: Ihlanganisa i-DNA, i-RNA, amaprotheni kanye nokuhlaziywa kokusebenzisana kwamamolecule ezinto eziphilayo, njengokuklama nokwenza ngcono ngokuqhubekayo ama-protein binder okungenzeka asetshenziswe ekuqondiseni i-IL-7Rα ekwelashweni kokuzivikela komzimba.
Ukuklama izinto zokwakha nezakhiwo zamakristalu: Ngokusekelwe kufomula yamakhemikhali, iqedela ngokuzenzakalelayo ukufanisa amaqembu esikhala kanye nokwakha ngendlela ehlelekile amapharamitha e-lattice nezixhumanisi zama-athomu, iguqule ukuhlolwa kwezinto ezintsha kusuka “ekuzameni nasephutheni ngokusekelwe kokuhlangenwe nakho” kuya “ekuqhutshweni imodeli”.
Ucwaningo lwezibalo nethiyori: Ibhekana nezinkinga eziyinkimbinkimbi zezibalo zokuncintisana, izibalo eziphakeme kanye nezinga le-PhD entrance exam, iqhubeke amahora amaningi icabanga futhi iqinisekisa, ikhiqize izimpendulo eziqinile futhi isekele ucwaningo lwethiyori.
Ukwakhiwa kobunjiniyela nokuzenzakalela kwe-agent ehlakaniphile: Ngomusho owodwa ikhiqiza imiphumela yobunjiniyela eyinkimbinkimbi, iqedele ngokuzenzakalelayo ukusesha kolwazi, ukuhlukanisa imisebenzi nokuhlela ukubambisana namathuluzi amaningi.