I-Artificial Analysis ishicilele umbhalo webhulogi ngo-Agasti 24, yamemezela ukubambisana ne-Liquid AI ekukhipheni isilinganiso sokuhlola ukusebenza kwe-AI kumaselula, kugxilwe ekusebenzeni kwamamodeli e-AI ku-Apple iPhone 17 Pro.
I-Artificial Analysis ibhekele uhlelo lokuhlola izinga lobuhlakani, kanti i-Liquid AI ibhekele uhlelo lokuhlola ukusebenza kokucabanga; lezi zinhlangano zombili zikhethe izivivinyo eziyisisekelo ezi-5:
BFCL (izinga lokusebenzisa imisebenzi lase-Berkeley)
IFBench (ukuhlolwa kokulandela imiyalelo)
AA-Omniscience (ukuhlolwa kolwazi nokuqonda)
GPQA Diamond (ukuhlolwa kwemibuzo nezimpendulo enzima kakhulu)
MATH-500 (ukuhlolwa kwezinkinga zezibalo).
Amamodeli ahlolwayo ngamamodeli ancishiswe aba ngu-4 bit futhi angadluli ku-8GB ngosayizi; izibonelo zifaka i-Gemma 4 E2B ne-LFM2-2.6B-Exp.
Ngemva kokukhawulela ubude bomongo ku-16K tokens, i-Nanbeige4.2-3B ne-LFM2.5-2.6B yizona ezithole amaphuzu aphakeme kakhulu ngokwesilinganiso ezivivinyweni eziyisisekelo. Lapho isikhathi sokuphendula sikhawulelwe kumzuzu ongu-1, i-LFM2.5-8B-A1B ibe sendaweni yokuqala, ilandelwa i-LFM2-2.6B-Exp, i-Gemma 4 E4B (Non-reasoning) kanye ne-Granite 4.1 8B.
I-Nanbeige4.2-3B yethulwe yi-Nanbeige Laboratory engaphansi kwe-BOSS Zhipin. Inezigidigidi ezingu-3 kuphela zamapharamitha angewona awokushumeka (amapharamitha aphelele acishe abe yizigidigidi ezingu-4), futhi isebenzisa ukwakheka kwe-Looped Transformer. Ngokusebenzisa kabusha izendlalelo ze-Transformer izikhathi ezimbili ngaphandle kokwandisa inani lamapharamitha, ithuthukisa ukujula kokubala namandla asebenzayo emodeli.
Ngezansi kunezithombe-skrini ezifanele:
