I-Thinking Machines Lab ikhiphe umbhalo webhulogi izolo (Julayi 30), yamemezela ukwethulwa kwemodeli ye-Inkling-Small, esebenza ngokucishe kufane ne-Inkling kodwa usayizi wayo wehliswe waba cishe ingxenye yesine.
I-Inkling ibizwa ngokuthi “imodeli ye-AI yomthombo ovulekile enamandla kunazo zonke eMelika”, isebenzisa ukwakheka kwe-mixture-of-experts (MoE), inamapharamitha angu-975B esewonke, amapharamitha asebenzayo angu-41B, kanye nomongo omude kunayo yonke onama-token angu-1M.
Imodeli ye-Inkling-Small esanda kwethulwa nayo isebenzisa ukwakheka kwe-mixture-of-experts (MoE), inamapharamitha angu-276B esewonke namapharamitha asebenzayo angu-12B, usayizi wayo ulingana cishe no-1/4 we-Inkling, futhi yaqeqeshwa ohlelweni lwe-NVIDIA GB300 NVL72.
Njenge-Inkling, i-Inkling-Small isekela ukucabanga komdabu ngomsindo nezithombe, amandla okucabanga aguquguqukayo, iwindi lomongo lama-token afinyelela ku-1M, futhi igcina ukusebenza okulinganiselayo ezinhlotsheni eziningi zokuhlolwa okuyisisekelo.
I-Thinking Machines isivule izisindo eziphelele ze-Inkling-Small, yaphinde yayihlanganisa nensizakalo yokwenza i-Tinker fine-tuning, kanti i-Tinker Playground inikeza ulwazi lokuxoxa ngombhalo, izithombe nomsindo.
I-Thinking Machines ithi i-Inkling-Small ithola ukusebenza okucishe kufane ne-Inkling isebenzisa amandla amancane okubala. Ezivivinyweni zokusebenzisa amathuluzi ama-agent, ukucabanga nokulandela imiyalelo, ukusebenza kahle kwe-Inkling-Small kuphakeme kune-Inkling, futhi ingancintisana namanye amamodeli anezinga elifanayo lezisindo.
Emisebenzini yokucabanga neyama-agent, i-Inkling-Small iyalingana noma idlule i-Inkling. Ekuhlolweni kwe-Humanity’s Last Exam, i-Inkling-Small ithole u-31.6%, odlula u-29.7% we-Inkling. I-Thinking Machines ithi ngaphansi kwesabelomali ngasinye sokucabanga, ijika lokubala ngesikhathi sokuhlolwa le-Inkling-Small lihlala liphakeme kune-Inkling.
