nmwdhj DeepSeek-V3 mftwH lmSdr lldhk lSTn`y fits `l~ Llama

DeepSeek tTlq nmwdhjan Dkhman b 671 mlyr m`ml wkf@ m`lj@ `ly@

lnmwdhj ytfwq `l~ lmnfsyn mftwHy lmSdr wyqtrb mn ‘d lnmdhj lmGlq@

tklf@ tdryb mnkhfD@ nsbyan blGt 5.57 mlywn dwlr mqrn@ bmy’t lmlyyn llnmdhj lmnfs@

‘`lnt shrk@ DeepSeek lSyny@ lnshy’@ `n Tlq nmwdhj ldhk lSTn`y ljdyd DeepSeek-V3, lmtH `br mnS@ Hugging Face. “mzyj” lnmwdhj “mzyj” lty, tufWl 671 mlyr Mml mm yDmn.

lmmyzt ltqny@ lry’ysy@:

• strtyjy@ mwzn@ Hml mtTwr@ ltwzy` l`ml byn wHdt lm`lj@
• Tqny@ Ltnbw’ MTP ttyH Sr@ Mlj@ 60 rmzan Fy lthny@
• Qdr@l mlj@syq nSyy ySl 128 ‘lf Rmz

Tdrb Lnmwdhjl 14.8, trylywn, rmz brmjy@w

l’d wlmqrnt:

• tfwq `l~ lnmdhj mftwH@ lmSdr mthl Llama-3.1-405B wQwen 2.5-72B
• ‘d mtmyz fy lkhtbrt lSyny@wlryDyt, m` tsjyl 90.2 fy khtbr Math-500
• is close to lnmdhj lmGlq@ mthl GPT-4 w Claude 3.5

lnmwdhj mtH Hlyan `br:

• GitHub brkhS@ MIT llkwd lmSdry
• wjh@ brmj@ ltTbyqt (API) llstkhdm ltjry b’s`r tnfsy@

ymthl hdh ltTwr khTw@ mhm@ fy tqlyS lfjw@ byn lnmdhj mftwH@ wmGlq@ lmSdr, mm ywfr khyrt mt`dd@ llshrkt wymn` Htkr tqnyt ldhk lSTn`y.

tb` `lm ltqny@ `l~