Beam Uyasebenza

I-Beam iyimodeli yokuqala enkulu yolimi yomthombo ovulekile evela kwa-Reflection AI, esebenzisa ukwakheka kwe-sparse MoE okunamapharamitha angu-501B, lapho kwenziwa kusebenze angu-23B kuphela kutokeni ngayinye. Isekela umongo wamathokheni ayisigidi futhi ukuqeqeshwa kwayo kwangaphambili kwasebenzisa amathokheni angu-23.8T. Inqubo yokuqeqesha i-Beam ingachazwa njengombukiso omkhulu wamandla ekhompyutha: ukuqeqeshwa kwangaphambili kuqedwe ngaphansi kwamasonto amane kusetshenziswa ama-GB300 angu-6,144, kwase kusetshenziswa ama-GB300 angu-10,500 ekuqeqeshweni kokuqinisa okuthathe amasonto amane futhi kwenza ama-rollout angaphezu kwezigidi eziyi-100.

Ukuthuthukiswa kwezinhlelo: Ifinyelela ezingeni elifanayo nele-GLM-5.2 emisebenzini yobunjiniyela besoftware futhi iyimodeli eyinhloko yemithwalo yokubhala amakhodi yamabhizinisi.
Ukusebenza kwama-agent: Inamandla okubiza amathuluzi, ukusebenzisa i-terminal nokusesha iwebhu, futhi ingaqedela imisebenzi enezinyathelo eziningi ngokuzimela, kuhlanganise nokufunda ukubiza amanye amamodeli amakhulu olimi nama-API e-OCR.
Ukucubungula imibhalo emide: Isekela umongo wamathokheni ayisigidi, ikwazi ukuthola nokuxhumanisa ulwazi olubalulekile kumakhosombe amakhulu kakhulu ekhodi nasemibhalweni.
Ukujula kokucabanga okulungisekayo: Umsebenzisi angashintsha kalula phakathi kwezimpendulo ezimfushane nokucabanga okude ngepharamitha ye-reasoning effort, ukuze ihambisane nemisebenzi ehlukene nesabelomali samandla ekhompyutha.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxoxa ngamaphrojekthi e-AI omthombo ovulekile
Ukwakhiwa kwe-sparse MoE: Imodeli inamapharamitha angu-501B esewonke, kodwa kwenziwa kusebenze angu-23B kuphela ngesikhathi ngasinye. Ukuqondisa ochwepheshe ngezinga elincane, kuhambisana nokulinganisa umthwalo okungancikile ekulahlekeni kosizo, kufinyelela ekusetshenzisweni kwabochwepheshe okucishe kulingane; umthwalo wochwepheshe omatasa kakhulu ungama-1.04× kuphela. Lokhu kuyisisekelo sezinga lehadiwe elenza i-Beam icubungule kahle.
Ukuqeqeshwa okukhulu ngezigaba ezimbili: Ukuqeqeshwa kwangaphambili kugaya amathokheni angu-23.8T ngaphansi kwamasonto amane kuma-GB300 angu-6,144, kuhambisana nomhleli owazi i-topology kanye nokutholwa kokonakala kwedatha buthule, futhi kufinyelela ku-goodput engu-92.3%; ngemva kwalokho kusetshenziswa ama-GB300 angu-10.5K ekuqeqeshweni kokuqinisa kwamasonto amane, okwenza ama-rollout angaphezu kwezigidi eziyi-100, kugcinwe ngokwesilinganiso ama-rollout angu-110,000 asebenzayo ngesikhathi esisodwa nama-sandbox angu-170,000 asebenzayo ngesikhathi esisodwa.
Uhlelo lwe-RL olusebenza ngokuphelele ngokungavumelani: Lusebenzisa ama-asynchronous policy gradients futhi lusekela ukusebenza ngokuhambisana kwe-rollout nokuqeqeshwa. Inkinga yokuguga kwe-policy ixazululwa ngokumaka ithokheni ngayinye ngenguqulo yemodeli eyayikhiqiza yona; noma isampula lisalele i-policy yamanje ngosuku olulodwa (izinguqulo ezingu-107), amanani ahlala ezinzile. Isisindo esisha sisatshalaliswa ngokwezigaba nge-RoCE siye kumaqoqo oku-inference ngemizuzwana engu-12 nje ngokwe-median.
Ukufunda ukucabanga okulawulwa ngobude: Ku-RL kufakwa inhlawulo yobude, ukuze imodeli iqale ifunde ukuxazulula izinkinga ngamathokheni ambalwa, bese yandisa ubude bokucabanga ngokwemvelo ngemva kokuthuthuka kwamakhono e-agentic. Lokhu kwakha ukulingana okuhle phakathi kwamakhono nokusetshenziswa kwamathokheni, okulungiswa ngumsebenzisi ngokwesidingo.

Kulesi sigaba, izisindo azikakhishwa; inguqulo yokubuka inikezwa abasebenzisi abathile kuphela. Vakashela ku-https://platform.reflection.ai/ ukuze ufake isicelo sokuyizama kusenesikhathi.

Ukusebenza kahle kokucabanga kuphezulu kakhulu: Ifinyelela ezingeni elifanayo lokucabanga ne-GLM-5.2 isebenzisa amandla e-inference aphansi ngokuphindwe ka-3–4, futhi yonga ngaphezu kwamamodeli e-2T+ afana ne-Qwen 3.8-Max, okugqamisa ukuminyana kobuhlakani.
Amapharamitha asebenzayo mancane kakhulu: Kumapharamitha angu-501B esewonke, kwenziwa kusebenze angu-23B kuphela kutokeni ngayinye, futhi ukwakheka kwe-sparse MoE kwehlisa kakhulu izindleko zokuyisebenzisa nezokucubungula.
Isikali se-RL asivamile embonini: Ama-GB300 angu-10.5K asetshenziswa amasonto amane ngokulandelana, kwenziwa ama-rollout angaphezu kwezigidi eziyi-100, okungenye yezinhlelo ezinkulu kakhulu zokuqeqesha i-RL ezake zenziwa elabhorethri evulekile kuze kube manje.
Ukuzinza kokuqeqeshwa kuqinile: Kube nokubuyiselwa emuva okuzenzakalelayo kancane izikhathi ezingu-9 kuphela kukho konke ukuqeqeshwa kwangaphambili, i-goodput yafinyelela ku-92.3%, futhi ijika lokuqeqeshwa alizange lingazinzi, okubonisa ukuvuthwa okuphezulu kobunjiniyela.
Ubude bokucabanga buyalawuleka: Umsebenzisi angalungisa ngokukhululekile ukujula kokucabanga ngepharamitha ye-reasoning effort, alinganise ukusebenza nezindleko ngokwesidingo.

Iwebhusayithi yephrojekthi: https://reflection.ai/blog/introducing-beam

Umsizi wokubhala amakhodi ozimele wamabhizinisi: Amabhizinisi angazakhela umsizi we-AI wokubhala amakhodi onezindleko eziphansi, ozisingathwa ngokuphelele ngaphakathi, ukuze idatha ingaphumi kunethiwekhi yangaphakathi.
Ukuhlelwa kabusha nokucwaningwa kwekhodi ezingeni lekhosombe: Ngewindi lomongo lamathokheni ayisigidi, i-Beam ingalayisha ikhosombe eliphakathi lonke ngesikhathi esisodwa, iqonde ubudlelwano phakathi kwamafayela futhi yenze ukuhlelwa kabusha okukhulu noma ucwaningo lokuphepha.
Ukugeleza komsebenzi kwama-agent okuzenzakalelayo: I-Beam iqeqeshelwe ngamandla nge-RL esekelwe kuma-rollout angaphezu kwezigidi eziyi-100, yafunda ukubiza amathuluzi, ukusebenzisa i-terminal nokusesha iwebhu, futhi ingaqhuba imisebenzi emide enezinyathelo eziningi efana nokuhlaziywa kwedatha nokuhlolwa kokusebenza kwezinhlelo.
Ukusatshalaliswa kwe-AI ezimele kahulumeni namabhizinisi: Njengemodeli enezisindo ezivulekile eyakhiwe ithimba laseMelika, i-Beam iyacwaningeka futhi ingasetshenziswa endaweni ehlukanisiwe, ihambisane nezidingo zokuthengwa okuhambisana nemithetho kukahulumeni, ezamandla nezokuvikela, ikakhulukazi kumamodeli angaveli ochungechungeni lokuhlinzeka lwaseChina.