Naive-N0.5-Flash Uyasebenza

I-Naive-N0.5-Flash iyimodeli yokuqala enkulu ye-MoE ekhishwe njenge-open source yi-NaiveAI, egxile ezinhlelweni, ekubhaleni amakhodi nasekuthuthukisweni kwe-AI. Yakhiwe phezu kwemodeli eyisisekelo i-MiMo-V2.5 yakwa-Xiaomi, inesakhiwo se-309B MoE, isebenzisa ukunaka okuyingxubevange okungasindi kwe-SWA+DSA, isekela i-context engu-1M ngokwemvelo ngaphandle kwezingqimba zokunaka okuphelele. Imodeli ithuthukiswe ngokubamba iqhaza okujulile kwe-AI, kanti uhlelo lokuqagela i-NaiveRT lufinyelela ku-2122 tokens/s ekudekhodweni komfudlana owodwa.

Ukuhlela nge-context ende: isekela i-context engu-1M token ngokwemvelo, futhi ingacubungula ama-code repository amade kakhulu nemisebenzi yobunjiniyela besofthiwe yesikhathi eside.
Usizo ocwaningweni nasekuthuthukisweni kwe-AI: ingafunda amaphepha, iphinde ikhiqize izindlela, iqhube izivivinyo, ihlaziye imiphumela, futhi ibambe iqhaza ngqo kuyo yonke inqubo yokuthuthukisa amamodeli.
Ukusheshisa ukuqagela: uhlelo lwe-NaiveRT oluhambisana nayo lusebenza ngo-50 tokens/s kumodi ejwayelekile, kanti kumodi eshesha kakhulu lufinyelela ku-2,122 tokens/s ekudekhodweni komfudlana owodwa.
Ucwaningo oluzimele: ngemva kokunikezwa umgomo, isabelomali samandla ekhompyutha kanye nephrothokholi yokuhlola, ingaqhubekisela phambili ucwaningo ngokuzenzakalela kusukela ekuqaleni kuye ekugcineni.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lephrojekthi ye-AI open source
Ukunaka okuyingxubevange okungasindi kwe-SWA–DSA: imodeli isekelwe kumodeli eyisisekelo i-MiMo-V2.5, futhi izingqimba zokunaka komhlaba wonke ezibiza kakhulu zithathelwe indawo ukunaka okungasindi kwe-DeepSeek. Lokhu kushintshaniswa nokunaka kwewindi elishelelayo ngesilinganiso esingu-5:1, ngaphandle kwanoma yiliphi ungqimba lokunaka okuphelele kuyo yonke inethiwekhi. I-SWA icubungula iwindi lendawo lama-token angu-128, kanti i-DSA isebenzisa i-indexer elula ukukala wonke umlando bese ikhetha ama-token angu-top-2048 ukuze acutshungulwe umgogodla.
Ukwakhiwa kwe-indexer elula: i-MLA yasekuqaleni ye-DSA ishintshwe yaba yi-GQA, kanti amakhanda emibuzo e-indexer ancishisiwe kusuka ku-64 aya ku-16. Lokhu kwehlisa isikhathi sendlela yokukhetha i-index ngo-44%, kuyilapho ukusebenza kugcinwa kunganciphi emisebenzini ye-Agent ene-context engu-1M.
Inqubo yokuqeqesha yezigaba ezintathu: okokuqala umgogodla uyamiswa ku-50B token, bese i-indexer iqondaniswa nokusatshalaliswa kokunaka okuphelele kusetshenziswa ukulahleka kwe-KL divergence; bese kushintshelwa ekunakeni okungasindi ukuze kuqhutshwe ukuqeqeshwa kwangaphambili kuma-token angu-3T, kujwayelaniswe indlela entsha yokukhetha ulwazi futhi kuqiniswe amakhono okuhlela nawocwaningo; ekugcineni kwenziwa i-supervised fine-tuning kuma-token angu-200B futhi kwehliswe izinga lokufunda. I-context engu-1M igcinwa kuyo yonke inqubo, kanti i-AI iqedela ngokuzenzakalela ukuhlaziywa kokubuyisa i-top-k nokulungiswa kokuzinza kwezinombolo.
Ukufana kokulandelana okuyingxube nokuthuthukiswa kwenkumbulo yevidiyo: ngenxa yokwehluka kokuthi i-SWA idinga iwindi lendawo kuphela, kanti i-DSA idinga wonke umlando, kusetshenziswa ukufana kokulandelana kwe-Halo ngokuhlukile; ohlangothini lwenkumbulo, ukulayishwa ngaphandle kwama-activation kwenziwa ezingeni lama-operator, futhi ama-index e-top-k agcinwa ngokusobala ukuze kugwenywe ukubala kabusha okungashintsha ukukhetha. Amakhadi angu-512 angakwazi ukuqeqesha ama-token angu-1T cishe ezinsukwini ezi-4.
Ukuthuthukiswa kokuqagela kwe-NaiveRT: ngokuhlanganisa ama-operator amakhulu, ukusebenzisa i-PDL ukuze ama-kernel aphezulu naphansi asebenze ngokugqagqana, kanye nokusebenzisa umjikelezo wokuqagela ngokushesha ngokuphelele ku-GPU, ukubambezeleka kokudekhodwa komjikelezo owodwa kwehliswa kusuka ku-12.3ms kuya ku-3.4ms; futhi kusekela ukusampula okunqunywa ngebhithi ngayinye ngokwembewu engahleliwe, okuqinisekisa ukuthi ukusampula ne-training logprob kuyavumelana ekuqeqeshweni kwe-RL. Konke ukuthuthukiswa kuhlolwa ngokwendlela yangempela yemodeli kusukela ekuqaleni kuye ekugcineni.

Qinisekisa ihardware: kudingeka i-NVIDIA GPU esekela i-FP8, isisindo semodeli sicishe sibe ngu-315GB, futhi kufanele kugcinwe enye inkumbulo yevidiyo eyengeziwe.
Thola izisindo: landa i-repository ye-NaiveAI/Naive-N0.5-Flash-FP8 ku-Hugging Face.
Faka okuncikile: sebenzisa pip install "transformers[torch,kernels]>=5.17.0".
Layisha imodeli: sebenzisa i-AutoTokenizer ne-AutoModelForCausalLM, kulayishwe nge-trust_remote_code=True kanye ne-device_map="auto".
Yakha okufakwayo: sebenzisa i-apply_chat_template ukudlulisa imilayezo yengxoxo nokukhiqiza ama-tensor okufakwayo.
Khiqiza imiphumela: shayela i-model.generate() ukuze ukhiqize impendulo, bese usebenzisa i-tokenizer.decode() ukuze u-decode okukhiphayo.
Setha amapharamitha: ezimweni ezijwayelekile kunconywa i-temperature=1.0 kanye ne-top_p=0.95.
Phinda ukukhiqizwa kokuqagela kokusebenza okuphezulu: uma kudingeka isivinini esingu-2,122 tokens/s, thola ikhodi ye-NaiveRT open source, indawo ye-Docker Compose kanye nezikripthi ze-benchmark ezihambisana nayo ukuze uzifake ngokwakho.

I-context ende kakhulu: isekela i-context engu-1M token ngokwemvelo, futhi ayinazo izingqimba zokunaka okuphelele kuyo yonke inethiwekhi, ngakho izindleko zokudekhodwa kokulandelana okude ziphansi kakhulu kunezakhiwo ezijwayelekile.
Isivinini sokuqagela esihamba phambili: uhlelo lwe-NaiveRT oluhambisana nayo lufinyelela ku-2,122 tokens/s ekudekhodweni komfudlana owodwa, kanti ukubambezeleka kokuqagela ngokushesha kwehliswe ngo-72.4% uma kuqhathaniswa ne-SGLang.
Ikhono eliqinile locwaningo nentuthuko: i-PaperBench 63.2, i-MLE-bench-30 73.7%, ne-NanoGPT SpeedRun 73.8s, okudlula amamodeli asezingeni elifanayo ezivivinyweni eziningi zocwaningo nentuthuko ye-AI.
Ucwaningo nentuthuko oluqondiswe yi-AI: imodeli ibambe iqhaza ekwakhiweni kwezakhiwo nasekuthuthukisweni kohlelo, okusebenzisa iphethini ethi “AI ithuthukisa i-AI”, lapho abantu bebeka khona imigomo bese i-AI yenza izimpinda zokuthuthukisa.
Amathuba ocwaningo oluzimele: ngemva kokunikezwa umgomo nesabelomali, ingaqedela ucwaningo ngokuzimela kusukela ekuqaleni kuye ekugcineni; ngemva kwamahora angu-400 okuqeqeshwa ikhiqize imodeli yomhlaba i-AutoWM futhi yafinyelela phezulu ku-WorldArena.

Iwebhusayithi esemthethweni yephrojekthi：https://naive.ai/en/research/
I-repository ye-GitHub：https://github.com/NaiveAI-Labs/Naive-N0.5-Flash
I-repository yemodeli ku-Hugging Face：https://huggingface.co/NaiveAI/Naive-N0.5-Flash

Umsizi wocwaningo nentuthuko ye-AI: usiza abacwaningi ukufunda amaphepha, ukuphinda bakhiqize izindlela, baqhube izivivinyo futhi bahlaziye imiphumela, njengasemisebenzini yocwaningo efana ne-PaperBench ne-MLE-bench.
Ubunjiniyela besofthiwe besikhathi eside: ngenxa ye-context engu-1M eyakhelwe ngaphakathi, ingacubungula ama-code repository amakhulu, iqonde amafayela ahlukahlukene, ilungise iziphazamisi futhi ikhiqize amakhodi ezingeni le-repository.
Imisebenzi ye-Agent ene-context ende: igcina wonke umlando womsebenzi ezimweni zokuxhumana ezinde, njengokusebenza ku-terminal nokushaya amathuluzi izikhathi eziningi.
Ingqalasizinda yokuqeqesha i-RL: ihambisana nesivinini esiphezulu sokudekhodwa komfudlana owodwa se-NaiveRT kanye nokusampula okunqunywa ngebhithi ngalinye, isheshise i-rollout ye-reinforcement learning yesikhathi eside futhi yehlise amasampuli asala emuva.