Faka kunye sebenzisa ubukrelekrele bokwenziwa kwindawo yakho Ukuba namagumbi okuncokola kwifowuni yakho okanye kwikhompyutha akusengokwabantu abathanda ubuchwepheshe kuphela okanye iinkampani ezinkulu. Namhlanje, nabani na angenza eyakhe "iChatGPT eyenziwe ekhaya" esebenzisa iitemplate ezivulelekileyo ezifana neLlama, DeepSeek, Phi, Gemma, okanye iMistral aze azisebenzise ngaphandle kweintanethi, idatha yakho ihlala ikwisixhobo sakho.
Ingcinga ilula: endaweni yokuthumela imibuzo yakho kwiiseva ze-OpenAI, Google, okanye Anthropic, ukhuphela imodeli ye-AI uze uyiqhube ngokwakho. Oko kubandakanya ezinye ukuzincama okunokwenzeka kunye nentuthuzeloKodwa ke, ufumana ubumfihlo, ulawulo olupheleleyo kwizicwangciso, kunye nenkululeko yokwenza ngokwezifiso i-AI ngendlela othanda ngayo. Makhe sijonge, ngokuzolileyo kodwa ngaphandle kokugxila kakhulu kwiinkcukacha, oko unokukwenza, oko ukufunayo, kunye nokuba zeziphi ii-application ezilungileyo zokuseta i-AI yakho yasekuhlaleni kwiselfowuni nakwiPC.
Kuthetha ukuthini ngokwenene "ukusebenzisa i-AI apha ekhaya", kwaye kutheni inokuba nomdla kuwe?
Xa uthetha ngayo sebenzisa i-AI apha ekhaya Sithetha ngokusebenzisa iimodeli ze-AI ngqo kwisixhobo sakho: imodeli, ubunzima bayo, kunye nokucubungula konke kwenziwa kwiPC yakho, kwilaptop, okanye kwisixhobo esiphathwayo, ngaphandle kokuxhomekeka kwiiseva zangaphandle. Oku kwahlukile kwiincedisi eziqhelekileyo ezifana neChatGPT, Gemini, Copilot, okanye Claude, ezisebenza efini kwaye zikunika iziphumo kuphela.
Ngokwenza njalo, nayiphi na isindululo, uxwebhu okanye idatha eyimfihlo Ulwazi olusebenzisayo aluze luphume kumatshini wakho. Akukho nto idlula kwiiseva zomntu wesithathu ngaphandle kokuba ukhetha ngokucacileyo ukuqhagamshela kwiimodeli zorhwebo nge-APIs. Kwiindlela ezininzi zokusetyenziswa kwemihla ngemihla (ukubuza imibuzo, ukubhala ii-imeyile, ukuvelisa ikhowudi, njl.njl.), abancedisi be-intanethi banele, kodwa ukuba uphatha ulwazi lwezonyango, lwezemali, lwezomthetho, okanye lwenkampani, ithuba lokuvuza kwelifu alivumelekanga.
Ngaphezu koko, ukusebenzisa i-AI kwindawo yakho kukuvumela ukuba uzame iimodeli zomthombo ovulekileyo Ngaphandle kwemiqathango eqhelekileyo: izihluzi zomxholo ezimbalwa, ukukwazi ukutshintsha indlela yokuziphatha, ukulungisa inkqubo yemiyalelo, okanye ukuyidibanisa nedatha yakho (ii-RAG, iiarhente, izixhobo zangaphandle, njl.njl.). Nangona kunjalo, ezi modeli zihlala zingenamandla kangako kuneenkampani ezinkulu zorhwebo kwaye zifuna izixhobo ezifanelekileyo ukuze zisebenze kakuhle.
Iingenelo kunye neengozi ze-AI esekwe kwilifu xa kuthelekiswa ne-AI esesitratweni
Abancedisi abasekelwe kwilifu baluncedo kakhulu kuba Abaxhomekekanga kumandla Kwisixhobo sakho: ungasebenzisa i-GPT-4, iGemini, okanye iClaude kwilaptop okanye kwisixhobo esiphathwayo esincinci ngaphandle kokukhathazeka nge-RAM okanye i-GPU. Ngokwesiqhelo zibonelela ngokufikelela okungcono kulwazi lwamva nje, ii-plugins, uphendlo lwewebhu, kunye namava omsebenzisi aphucukileyo.
Icala elisezantsi lelo Yonke into oyibhalayo iyarekhodwa ngaxa lithile kwiseva, ubuncinane okwethutyana. Kubekho iziganeko apho iincoko zangaphakathi okanye idatha ziye zavezwa ngenxa yokwaphulwa kokhuseleko. Ukuba udala amabali, iingcamango zobuchule, okanye iindlela zokuhamba, umngcipheko uphantsi kakhulu; kodwa ukuba ufaka amagama ayimfihlo, iinombolo zekhadi letyala, iirekhodi zonyango, okanye idatha yabathengi, izinto ziyatshintsha.
Nge-AI yasekuhlaleni, ukucutshungulwa kwenziwa ngokupheleleyo kwikhompyutha yakho okanye kwisixhobo esiphathwayo. Akukho mntu wesithathu ubona idatha, kwaye unako sebenza ungaxhunyiwe kwi-intanethi ngokupheleleyoIxabiso ekufuneka ulihlawule kukuba udinga isixhobo esine-RAM eyaneleyo, i-VRAM, kunye nendawo yokugcina, kwaye kufuneka ujongane nokufakelwa, ukhuphelo lweemodeli, kunye nokuseta kokuqala. Ngaphezu koko, ngokuzenzekelayo, ezi modeli azinawo ukufikelela kwi-intanethi okanye amandla okukhangela idatha ngexesha langempela ngaphandle kokuba uzilungiselele ngokucacileyo.
Eyona ndlela ilungeleleneyo idla ngokuhlanganisa omabini la mazwe: imifuziselo yendawo malunga nomxholo obuthathaka kunye nemisebenzi yangaphakathi, kunye iimodeli zelifu Kwimibuzo ngokubanzi, uphendlo lwewebhu, okanye umsebenzi ofuna umgangatho ophezulu wokubhala umbhalo, umfanekiso, okanye ukuveliswa kwekhowudi.
Zeziphi izixhobo ozidingayo ukuze usebenzise i-AI kwindawo yakho?
Awudingi ikhompyutha enkulu ukuze uqalise, kodwa kuyanceda ukwazi ukuba yintoni Izixhobo zenza umahluko Xa usebenzisa iimodeli zasekuhlaleni, iimfuno ziyahluka ngokuxhomekeke kuhlobo lwemodeli (umbhalo, umfanekiso, i-multimodal) kunye nobukhulu bayo, kodwa kukho amanqaku afanayo.
Kwiimodeli zolwimi ezinkulu (LLM) isitshixo sikwi Imemori yevidiyo kunye ne-RAMI-RAM igqiba ukuba imodeli iyahambelana na kwaye zingaphi iinkqubo ezihambelanayo ezinokuqhutywa, ngelixa i-VRAM ye-GPU inefuthe kwisantya sokwenziwa. Nge-RAM okanye i-VRAM encinci kakhulu, i-AI yendawo isebenza ngesantya se-snail: amagama ali-1-2 ngomzuzwana, anele ukulingwa kodwa acotha kakhulu ukusetyenziswa kwemihla ngemihla.
Njengendawo efanelekileyo yokuqala kwikhompyutha yedesktop okanye yelaptop, kucetyiswa ngokubanzi ukuba uqale kwi 16 GB ye RAMI-CPU yanamhlanje (umzekelo, i-2017 Core i7 enenkxaso ye-AVX2 iya kwanela), kunye ne-GPU enobuncinci i-4 GB ye-VRAM. Ngaphantsi koko ungaqhuba iimodeli ezincinci, kodwa kuya kufuneka ukhethe iinguqulelo ezixineneyo kakhulu kwaye wamkele amaxesha okuphendula acothayo.
KwiiMacs, iitships zeApple Silicon (M1, M2 kunye nezamva) ziluncedo kuba imemori edibeneyo isebenza njengeVRAMLe nkqubo ingasebenzisa ukuya kuthi ga kwi-75% ye-RAM njengememori yevidiyo, nto leyo evumela ukuba ikwazi ukusingatha iimodeli ezinkulu kwi-MacBook Pro okanye kwi-Mac Studio, ingakumbi kwiinguqulelo ze-Max okanye ze-Ultra ezine-RAM eninzi.
Iindaba ezimnandi zezokuba zikho iimodeli ezilinganisiweyo nezilula ezisebenza kakuhle nakwizixhobo ezindala, zilahla ukuchaneka okuthile kodwa zigcina umgangatho ongaphezulu kowamkelekileyo kwimisebenzi yombhalo, izishwankathelo, okanye izixhobo ezincinci zokubhala iikhowudi.
Iingcamango ezisisiseko: iimodeli, iiparameter, umxholo, kunye nobungakanani
I-LLM (iModeli yoLwimi olukhulu) yiyo "Ingqondo" yomncedisi ye-AI. Yifayile enkulu oyikhuphelayo equlethe ubunzima obuqeqeshiweyo: amanani amele ulwazi kunye nemithetho esetyenziswa yimodeli ukuvelisa umbhalo. Iimodeli ezifana neLlama 2, iMistral, iGemma, iPhi, okanye iDeepSeek ziziseko apho iinguqulelo ezikhethekileyo zakhiwe khona kamva.
Uninzi lweemodeli oza kuzifumana zezi "Iingoma ezintle"Ezi ziinguqulelo ezilungiselelwe imisebenzi ethile (incoko, ucwangciso, izibalo, ukudlala indima, ukuguqulela, njl.njl.). Amagama anjengeWizard, Vicuna, Nous-Hermes, CodeLlama, WizardMath, okanye Orca Mini abonisa iindlela ezahlukeneyo zoqeqesho olongezelelweyo. Amaxesha amaninzi, ezi zidityaniswe (umzekelo, imodeli yeWizard-Vicuna) ukuzama ukufumana okona kulungileyo kwiindlela ezahlukeneyo.
Ubungakanani bemodeli buchazwa ku iibhiliyoni zeeparameter (3B, 7B, 13B, 34B, 70B…). Okukhona iiparameter zininzi, kokukhona imodeli ikwazi ngakumbi, nangona ukusetyenziswa kwememori nako kuyanda. I-70B inokusebenza phantse njengomntu kwincoko ende, ngelixa i-3B inobunzima ukuba incoko iba nzima. Okubangela umdla kukuba, abasebenzisi abaninzi bacinga ukuba iimodeli ze-13B ezilungisiweyo kakuhle zibonelela ngomlinganiselo ogqwesileyo wokusebenza nokusebenza ngokubanzi.
Enye ingcamango engundoqo umxholo"Ifestile" yememori esetyenziswa yimodeli ukuvelisa impendulo nganye. Iimodeli zokuqala zeLlama zaziphethe amakhadi omxholo angama-2048 (malunga namagama ali-1500), iLlama 2 idla ngokunyusa oku kuye kuma-4096 nangaphezulu, kwaye iimodeli zanamhlanje ziyandisa le festile nangakumbi. Okukhona umxholo umkhulu, kokukhona imbali yencoko, imiyalelo, kunye namaxwebhu onokuwathumela ngaxeshanye ngaphambi kokuba imodeli "ilibale" ulwazi lwangaphambili.
Okokugqibela, kukho ubungakananiLe yindlela evumela ezi modeli zinkulu ukuba zingene kwiikhompyutha eziqhelekileyo. Ubunzima beemodeli ekuqaleni bugcinwa ngocoselelo oluphezulu (umzekelo, iibhithi ezili-16), kodwa zinokuncitshiswa zibe yibhithi ezi-8, ezi-4, okanye ezi-2, nto leyo enciphisa kakhulu ubungakanani befayile kunye nememori efunekayo, ngexabiso lokulahleka kokuchaneka.
Enyanisweni, imodeli enkulu, elinganiswe kakhulu (umz., i-34B kwiibhithi ezi-3) ingenza ngcono kunemodeli encinci enobunzima obuchanekileyo, kuba inani leeparameters linobunzima obungaphezulu Ayikuko ukuchaneka okuluhlaza. Umdlalo ufumana indibaniselwano enkulu ehambelana ne-VRAM yakho kwaye iphendule ngesantya esamkelekileyo.
Apho ungazifumana khona iimodeli ze-AI kunye nokuba zeziphi iifomathi ezikhoyo
Indawo ephambili yokugcina iireferensi zokukhuphela iimodeli ezivulekileyo yile Ukujongana nobusoIcandelo layo leemodeli likuvumela ukuba uhluze ngobukhulu, uhlobo lomsebenzi, ilayisenisi, ubungakanani, kunye nezinye izinto ezininzi eziguquguqukayo. Ukusuka apho ungakhuphela iimodeli kwiLlama, Mistral, Gemma, DeepSeek, Phi, kunye nezinye ezininzi eziqeqeshwe luluntu.
Xa ujonga iimodeli uza kubona ezininzi iifomati zefayileIifomathi ezixhaphakileyo namhlanje yiGGUF, GPTQ, kunye ne-exl2, ukongeza kwiifomathi ezindala ezifana neGGML. Nganye yenzelwe uhlobo olwahlukileyo lokusebenza (iCPU+GPU, iGPU emsulwa, iilayibrari ezahlukeneyo). Uninzi lwezicelo zanamhlanje zisebenza kakuhle kakhulu kwiimodeli zeGGUF, ezilandela iGGML endala kwaye zidibanisa iCPU kunye neGPU ngokufanelekileyo.
Kukwakho neemodeli esele zilinganiswe ngokwee-precisions ezahlukeneyo (q2, q3, q4, q5, q6, q8), ngamanye amaxesha zinezimamva ezifana ne-K_S, K_M, okanye i-K_L ezibonisa ii-quantization variants. Njengomgaqo jikelele, kudla ngokuba ngcono ukusebenzisa imodeli enkulu enomlinganiselo ophakathi encinci enobunono obuphezulu ukuba izixhobo zakho ziyakuvumela.
Ukuba imodeli onomdla kuyo ayikho kwi-quantization oyifunayo, ungayilinganisa ngokwakho usebenzisa izixhobo ezifana GPTQ okanye ezinye izinto ezikhethekileyo, nangona okokuqala kudla ngokuba lula ukukhetha enye yeemodeli ezininzi esele zenziwe.
Ii-apps ezilungileyo zokuba ne-AI yasekuhlaleni kwiselula yakho
Ukuba ufuna ukuthwala umncedisi wakho we-AI epokothweni yakho, kukho ii-apps ezininzi ezikuvumela ukuba uzikhuphele kwaye uzisebenzise. Iimodeli ze-LLM ngqo kwiselula yakhoOku kufumaneka kwi-Android nakwi-iOS. Khumbula ukuba ezi modeli zithatha indawo eninzi, ngoko ke licebo elihle ukujonga indawo yakho yokugcina izinto ekhoyo ngaphambi kokuba uqale.
PocketPal AI Yenye yezona apps ziphambili zeselfowuni ze-AI yasekuhlaleni. Isimahla, ivulelekile, kwaye iyafumaneka kwi-Android nakwi-iOS. Eyona nzuzo yayo inkulu kukudibana kwayo ngqo ne-Hugging Face, ekuvumela ukuba ujonge kwaye ukhuphele iimodeli ngqo kwifowuni yakho, ngaphandle kokujongana nokukhuphela ngesandla okanye iindlela zefayile ezintsonkothileyo.
Ukusuka kwiPocketPal AI ungakhetha phakathi iimodeli ezininzi ezahlukeneyoUngazilungisa iiparameter ezisisiseko uze uthathe isigqibo sokuba uza kubeka phambili na isantya okanye umgangatho. Uyilo lugxile ekubeni lula: ujongano lwencoko olucocekileyo kunye nolawulo olulula lwemodeli yesixhobo esiphathwayo.
Kwi-Android, Incoko ye-MNN Igqame ngokuba yenye yezona ndlela zikhawulezayo kwaye inenkxaso epheleleyo ye-multimodal. Oku kuthetha ukuba ungathumela umbhalo, imifanekiso, okanye i-audio njengenxalenye yesicelo, usebenzisa iimodeli ezahlukeneyo zokubona, umbhalo, okanye ukuvelisa imifanekiso. Usetyenziso luquka ikhathalogu yangaphakathi yeemodeli kwaye lwenza kube lula kakhulu ukukhuphela nokufaka kwindawo yakho.
Kubasebenzisi be-iOS abafuna into ephucukileyo ngakumbi, I-LLM yabucala Yindlela yokuhlawula ekumgangatho ophezulu, ehlawulelwa kube kanye (malunga ne-$5). Iquka iimodeli ezingaphezu kwama-60 ezikhethwe kakuhle kwaye isebenzisa ubungakanani obuphambili ukuphucula ukusebenza kwezixhobo ze-Apple. Idibana neSiri kunye ne-Apple Shortcuts kwaye ingasetyenziswa kwi-iPhone, iPad, kunye ne-Mac, kunye neendlela zokwabelana ngokuthenga ngeFamily Sharing.
Unayo nakwi-Android Google AI Edge GalleryI-app kaGoogle eyilelwe ukuvavanya nokulawula iimodeli ze-AI kwisixhobo sakho kwimisebenzi efana nokuhlelwa kwemifanekiso kunye nokubuza imibuzo, ukuguqulelwa kwesandi, kunye nokuncokola. Yinkqubo evulelekileyo, kodwa isaphuhliswa, ngoko ke kuqhelekile ukudibana neempazamo okanye iimpawu ezingagqitywanga.
Kwinkqubo yendalo ye-Apple, I-AI yasekuhlaleni Inika amava acwebezelayo kwaye ilungiselelwe iiprosesa zeApple Silicon. Ivumela ukusetyenziswa kolwimi oluvulelekileyo kunye neemodeli zokubona, inemo yelizwi yasekuhlaleni, kwaye inikezela ngokudibanisa neSiri kunye neeShortcuts. Ingcinga kukuphinda ivakale njenge-app yohlobo lweChatGPT kodwa yonke into isebenza ngaphandle kweintanethi kwaye ngaphandle kokuxhomekeka kwiiseva zangaphandle.
Uza kuyifumana nakwi-Android. Nantoni naLLM kwinguqulelo yeselula. Le app ayijolisi kakhulu ekubeni nekhathalogu enkulu kodwa igxila kakhulu ekunikezeni iimodeli ezimbalwa ezikhawulezayo nezilungiselelwe kakuhle Kwizixhobo zeselfowuni. Ibandakanya imo yearhente ekuvumela ukuba wenze izinto ezifana nokufunda amaphepha ewebhu, ukwenza uphendlo, ukusebenzisana nezinye ii-apps, okanye ukusebenzisa indawo okuyo. Ukuba ufuna amandla angakumbi, ungaqhagamshela kwiimodeli zelifu zorhwebo, ngexabiso lobumfihlo obuthile.
Olunye ukhetho kwi-Android kukuba I-SmolChatYenzelwe ukukhuphela nokusebenzisa iimodeli ze-AI ezidumileyo zasekuhlaleni nge-interface elungiselelwe i-Android, inikezela ngeendlela ezininzi zokwenza ngokwezifiso kwaye ikuvumela ukuba unamathisele iincoko ozithandayo njengeendlela ezimfutshane kwisikrini sakho sasekhaya.
Izicelo zokufaka i-AI yasekuhlaleni kwikhompyuter yakho
Kwiikhompyutha zedesktop, ukhetho lubanzi nangakumbi; kukho ii-apps zedesktop ezinokukhangela kunye ne-AI Baqala ukudibanisa iimpawu ezinamandla. Ezi ziqala kwizixhobo ezilula kakhulu ezibonisa kuphela ifestile yencoko ukuya kumaqonga aphambili aneearhente, izihlanganisi kwiinkonzo zangaphandle, kunye neeseva zasekuhlaleni ezihambelana nee-API zesitayile se-OpenAI.
Musa Iye yaba yenye yezona ndlela zidumileyo zokusebenzisa i-AI yasekuhlaleni kwiikhompyutha. Isimahla, ivulelekile, kwaye iyafumaneka kwiWindows, macOS, kunye neGNU/Linux. Ayinawo ujongano oluntsonkothileyo lwemizobo; isebenza ngomgca womyalelo, nto leyo eyenza ibe lula kakhulu kwaye kube lula ukuyenza ngokuzenzekelayo.
Nge-Ollama ungafaka iimodeli ezifana DeepSeek, Llama, Phi, Gemma, Mistral, Qwen kunye nabanye, kwiinguqulelo ezahlukeneyo ngokuxhomekeke kwinani leeparameter. Umzekelo, i-DeepSeek-R1 version 7B ithatha iigigabytes ezimbalwa kwaye ingasebenza kwiikhompyutha ezine-8 GB ye-RAM, ngelixa iinguqulelo ezinkulu ezinamakhulu eebhiliyoni zeeparameter zifuna amashumi okanye amakhulu eegigabytes ze-RAM kunye nee-GPU ezininzi, into egcinelwe iindawo zobungcali.
Ukufaka iimodeli e-Ollama kusekelwe kwimiyalelo elula kakhulu, umzekelo uollama baleka deepseek-r1:8bLe nkqubo ngokwayo ilawula ukukhuphela imodeli, ukubonisa inkqubela phambili, nokuyilungiselela ukuncokola kwi-terminal okanye ukusetyenziswa njengenkonzo yasekuhlaleni zezinye iinkqubo. Ukuba ikhadi lakho lemizobo line-VRAM eyaneleyo, ukusebenza kokuvelisa kunokuba kuhle kakhulu, nokuba kukho iimodeli eziphakathi.
Ukuba ukhetha into ene-graphical interface, Isitudiyo se-LM Ibonelela ngesicelo esidibeneyo sokukhangela, ukukhuphela, kunye nokusebenzisa iimodeli ze-AI. Sisixhobo esivulekileyo kwaye sineenguqulelo zeWindows, macOS, kunye neLinux. Usebenzisa injini yayo yokukhangela edibeneyo, ungafumana iimodeli zeHugging Face, ufake izihluzi, uzikhuphele, kwaye uziqalise ngqo kwifestile yencoko.
I-LM Studio ikuvumela ukuba usebenzise iimodeli kwi-interface yazo okanye uziveze njenge iseva yendawo iyahambelana ne-OpenAI APIOku kuvumela izixhobo ezininzi ezenzelwe iChatGPT (iiklayenti zencoko, ii-plugins, ukuhlanganiswa okwenziwe ngokwezifiso) ukuba zisebenze nemodeli yakho yendawo ngokutshintsha nje i-API URL. Ikwavumela ukusebenza ngamaxwebhu asekuhlaleni, ukushwankathela, ukuguqulela, kunye nokwenza eminye imisebenzi yombhalo enzima ngaphandle kwe-intanethi.
Esinye isixhobo esinamandla kakhulu Nantoni naLLM kwinguqulelo yayo yedesktop. Ivulelekile kwaye ilungiselelwe ukuyifaka indawo yokusebenza epheleleyoIkuvumela ukuba uthethe ngamaxwebhu, usebenzise iiarhente ze-AI ukuze zenzele imisebenzi, uqhagamshelane neemodeli zasekuhlaleni, kunye nababoneleli belifu (i-OpenAI, i-Azure, kunye nabanye) ukuba uyafuna. Inoyilo oluguquguqukayo olunezixhobo ezininzi kunye nokugxila kakhulu kubumfihlo kunye nokwenza ngokwezifiso.
GPT4All Sisisombululo esidumileyo sedesktop. Sivulelekile kwaye sinokusebenza sisebenzisa iCPU kuphela okanye ngokusebenzisa iGPU xa ikhona. Ivumela ukufakwa kweemodeli zeelwimi ezahlukeneyo ukuya kuthi ga kwiwaka, kuquka iDeepSeek, iLLaMA, iMistral, iNous-Hermes, nezinye ezininzi. Ineenguqulelo zeWindows (kubandakanya i-ARM), iMacOS, kunye neUbuntu.
Nangona isicelo esipheleleyo sihlawulwe, sihlala sibonelela Inguqulelo yasimahla eneempawu ezilinganiselweyo Yanele ukusetyenziswa lula okanye uvavanyo. Inzuzo yayo ephambili kukuba ijoliswe kubasebenzisi abangafuni ngxaki yokufakelwa ngesandla: uyayikhuphela, ukhethe imodeli kwikhathalogu, kwaye ugqibile.
Jan Yinkqubo enye evulelekileyo enezigidi zokukhuphela ezikuvumela ukuba usebenzise iimodeli ezivulelekileyo (iLlama, iGemma, iMistral, njl.njl.) kwaye uqhagamshele kwiinkonzo zangaphandle ezifana ne-OpenAI okanye i-Anthropic. Lonke ugcino lwedatha lwenziwa apha ekhaya, kwaye lubonelela ngeenguqulelo zeWindows, macOS, kunye neGNU/Linux, ngenkxaso yeNVIDIA (CUDA), AMD (Vulkan), kunye nee-Intel Arc GPU.
UJan ukwabandakanya inkqubo yolwandiso kunye nezihlanganisi zokusebenza nazo I-Gmail, i-Amazon, i-Google, i-YouTube, i-Google Drive kunye nezinye iinkonzo, kunye nenkqubo yememori yasekuhlaleni esebenza kwisixhobo sakho. Lukhetho olufanelekileyo njenge "ziko lomyalelo" kwii-AI zakho, ezidibanisa izakhono zasekuhlaleni nezelifu.
Kubasebenzisi abaphambili, I-Msty Studio (ngamanye amaxesha ibizwa ngokuba yiMsty Studio) inikezela ngomnye wamava atyebileyo kakhulu. Ixhasa iimodeli zasekuhlaleni nge-Ollama, llama.cpp, kunye ne-MLX, kwaye inokunxibelelana nababoneleli bamafu ukuze basebenzise iimodeli zoshishino. Ivumela ukuhlanganiswa nee-API, izixhobo ze-MCP, ii-knowledge stacks, kunye nokudalwa kwe ukuqukuqela komsebenzi oqhelekileyosoloko ubeka phambili ukuba idatha eyimfihlo ayiphumi kwindawo okuyo.
Ukuba unomdla wokufikelela kwinqanaba eliphantsi, umnxeba.cpp Yinkqubo encinci, evulelekileyo eyenzelwe ukuqhuba iimodeli zeMeta ezisekelwe kwi-LLaMA apha ekhaya. Isebenza kwii-CPU kunye nee-GPU kwaye sisiseko apho kwakhelwe khona ezinye izixhobo ezininzi. Kunzima ngakumbi ukuyisebenzisa, kodwa isebenza kakuhle kwaye iguquguquka—ifanelekile ukuba ufuna ukufunda indlela esebenza ngayo ngaphakathi.
Esinye isiqwenga esinomdla ngu Ifayile yeLlamafayileI-Llama.cpp yiprojekthi yeMozilla Builders edibanisa i-llama.cpp kunye neCosmopolitan Libc ukupakisha iimodeli ze-AI njengeefayile ezizimeleyo ezisebenzisekayo. Oku kwenza kube lula ukusasaza iimodeli ezinokusebenza kwiWindows, Linux, macOS, okanye BSD ngokuvula ifayile enye, ngaphandle kofakelo oluntsonkothileyo.
Ukongeza, izixhobo ezikhethekileyo ziyavela ezinje Incoko yeNVIDIA neRTXYi-chatbot yasekuhlaleni eyenzelwe ii-PC zeWindows ezine-16 GB ye-RAM kunye ne-RTX 30 okanye 40 series GPU ene-8 GB ye-VRAM ubuncinci. Ingashwankathela iividiyo ze-YouTube, icwangcise iiseti zamaxwebhu, kwaye yenze eminye imisebenzi ngokusebenzisa iimodeli ezifana neMistral kunye neLlama 2. Yi-beta enzima (malunga ne-40 GB) kwaye inobunzima ukuyifaka, kodwa ibonisa apho i-GPU-accelerated local assistant ecosystem iya khona.
Indlela yokucwangcisa umsebenzi osebenzayo kunye ne-AI yendawo
Ukuze ufumane okuninzi ngokwenene kwi-AI yasekuhlaleni, kufanelekile ukucingisisa imisebenzi ethilehayi nje kuphela “ekuncokoleni nomatshini”. Umzekelo oqhelekileyo kukucubungula amaxwebhu ayimfihlo: iingxelo zonyango, iikhontrakthi, iingxelo zemali okanye amaxwebhu enkampani yangaphakathi.
Ngemodeli yasekuhlaleni, ungalayisha iiPDF okanye imifanekiso eskeniweyo, uyibhale phantsi, ufumane izishwankathelo ezilungiselelwe umgangatho wolwazi lomntu ofumana ulwazi (ingcali, isigulana, umthengi, njl.njl.), uvelise ii-imeyile okanye iingxelo ezibhaliweyo, kwaye uziguqulele kwezinye iilwimi—konke ngaphandle kokuba loo datha iphume kwikhompyutha yakho. Izixhobo ezifana neLM Studio, AnythingLLM, okanye uJan zikuvumela ukuba udibanise ukufundwa kwamaxwebhu kunye nencoko, nto leyo eyenza kube lula kakhulu le misebenzi.
Kwicala lobuchule, i-AI yasekuhlaleni inokusingatha yenza imifanekiso, phinda utolike izigcawuYenza iindlela zakudala zibe zezona zintsha kwaye ude wenze izinto ezisisiseko ziphile usebenzisa imiyalelo yombhalo. Amaqonga afana neComfy (asetyenziswa kakhulu ekuveliseni imifanekiso kunye neendlela zokusebenza ezibonakalayo) akuvumela ukuba wakhe imibhobho yokudala enzima, ugcine idatha kunye nemifanekiso kumatshini wakho kwaye uphephe imiba yobunini bolwazi enxulumene neenkonzo ze-intanethi.
Enye indlela enamandla yokusebenzisa kukucwangcisa: iimodeli ezifana neCode Llama okanye iinguqulelo ezikhethekileyo zinokukunceda ukuba bhala, uphonononge kwaye uchaze ikhowudi ngaphandle kokuba ii-repository zakho zabucala okanye iiprojekthi zabathengi ziye kwiiseva zangaphandle. Zidibene nabahleli okanye ii-IDE nge-API zasekuhlaleni ezihambelana ne-OpenAI, zinamandla amakhulu kwimisebenzi yophuhliso.
Eyona nto ibalulekileyo kukukhetha isixhobo esiphakathi esifanelekileyo (u-Ollama, i-LM Studio, i-AnythingLLM, uJan, i-GPT4All, njl.njl.), iimodeli ezithile zomsebenzi ngamnye, uze emva koko, wandise kancinci kancinci ngeearhente zakho, izihlanganisi, okanye izikripthi ngokweemfuno zakho zokwenyani.
Ekugqibeleni, ukusebenzisa i-AI kwindawo yakho kukuvumela ukuba ube ne umncedisi oguquguqukayo nowabucala Iyahambelana nezixhobo zakho: kwiimashini ezinamandla ungaqhuba iimodeli ezinkulu phantse kwinqanaba leenkonzo zorhwebo, nakwizixhobo ezincinci, nangona usenokuba ungakhawulezi kangako, uya kuqhubeka nokugcina ubumfihlo kunye nolawulo lwedatha yakho, ngaphandle kokuxhomekeka kubukho okanye utshintsho kwimigaqo-nkqubo yamaqonga amakhulu.
