I-Soofi Consortium Ikhipha I-Soofi S 30B-A3B: Imodeli Yesisekelo Ye-Hybrid Mamba-Transformer MoE YesiJalimane NesiNgisi

Inhlangano yocwaningo yaseJalimane ishicilele umbiko wokuziqeqeshela I-Soofi S 30B-A3B. Kuyimodeli eyisisekelo evulekile yesiJalimane nesiNgisi. Ukuqeqeshwa kuphele kwaze kwaba sekugcineni kuDeutsche Telekom's Industrial AI Cloud eMunich. Izisindo zokubuka kuqala ziku-Hugging Face. Kuhle ukuqaphela ukuthi phakathi kwamanye amamodeli ayisisekelo ahlolwe ngokugcwele, i-Soofi S irekhoda amaphuzu aphezulu kakhulu esiNgisi nesiJalimane.
Yini i-Soofi S 30B-A3B?
I-Soofi S iyimodeli yesisekelo ye-Mixture-of-Experts (i-MoE) eyihybrid Mamba Transformer. Ihlanganisa ~31.6B amapharamitha futhi ivule ~3.2B ithokheni ngayinye. Njengemodeli eyisisekelo, ayinakho ukushuna kweziyalezo, ukuqondanisa, noma ukushuna kokuphepha. I-KI Bundesverband iqondisa umfelandawonye, oxhaswe uMnyango Wezindaba Zomnotho Nezamandla waseJalimane. Abahlanganyeli bahlanganisa uFraunhofer IAIS, DFKI, TU Darmstadt, ellamind, kanye neMeantix Momentum.
Indlela i-architecture esebenza ngayo?
Isimangalo sokusebenza kahle siqala ngesitaki sesendlalelo. Inethiwekhi ibamba izendlalelo ezingama-52. Lezo izendlalelo ezihlanganisa ukulandelana kwe-Mamba-2 ezingu-23, izendlalelo ze-MoE eziyimbudumbudu ezingama-23, nezingu-6 zeGrouped-Query Attention (GQA). Yilezo zingqimba ze-GQA eziyisi-6 kuphela ezigcina inqolobane ye-KV. Isendlalelo ngasinye se-MoE siphethe ochwepheshe abathutha imizila abangu-128, senza kusebenze abangu-6 ngethokheni ngayinye, futhi sengeza ochwepheshe abahlanganyelwe abangu-2. Eminye imininingwane: ubukhulu bemodeli 2688, i-ReLU eyisikwele, i-RMSNorm, futhi akukho okushumekiwe kwezikhundla.
I-Soofi S yamukela idizayini yesithenjwa ye-Nemotron 3 Nano ngaphandle kokuguqulwa. Ithimba labacwaningi linikeza izizathu ezintathu zalokho kukhetha. Lokho wukuthunyelwa ezitakini ezifana ne-vLLM, ukusebenza kahle, nokulawulwa kwesayensi. Ngenxa yokuthi umgogodla ulungisiwe, i-Nemotron 3 Nano iba isisekelo esifana nesakhiwo. Iresiphi yedatha iyona kuphela ingxenye ehambayo.



