Kufarira kuita AI modhi pa-nzvimbo, pasina kuvimba negore, yakaputika zvichitevera kubuda kweDeepSeek. Kana iwe uine hanya nezve kuvanzika, kukurumidza, kana kuwanikwa pasina Indaneti, kuvaka yako yekuzvarwa AI muMac Mini ipfungwa huru: data rako rinogara kumba, hapana mari yemwedzi uye kushanda nayo Apple silicon Zvinoshamisa kune compact timu.
Pamusoro pekuvanzika, pane imwe mukana unoshanda: unogona kuyera kuita mukati zviratidzo pasekondi uye tarisa pachako kugona komputa kweMac yako. Uye chinhu chakanakisa ndechekuti maitiro ekuisa DeepSeek kushanda anowanikwa zvakanyanya nekuda kwezvishandiso zvakaita LM Studio (dzimwe nguva inonzi LLM Studio) uye Ollama, iyo inokutendera kuti utore, gadzirisa uye utaure nemuenzaniso nekungodzvanya zvishoma kana mirairo.
Chii chinonzi DeepSeek uye nei zvakakodzera kumisikidza munharaunda?
DeepSeek yakazunungusa nyika nezvayo mhando / mutengo reshiyo uye nzira yayo yakavhurikaKune maviri makuru anonaka ari mukutamba: iyo V3 yakatevedzana yekushandiswa kwewebhu, uye iyo R1 bazi, ine yakanyatso tarisisa kufunga, iyo inonyanya kunakidza kumhanya pakombuta yako mushanduro. distilledIdzi vhezheni dzakaderedzwa muhukuru kuti dzifambe zvakanaka pamidziyo yevatengi, uku ichichengeta yakawanda yekugona kwekutanga.
Chikumbiro chine musoro kana ukaisa pamberi kuvanzika uye kutongaKusiyana neyegore-yakavakirwa chatbot, hapana data inorodha kune ekunze maseva. Muzviitiko zvine kubatana kusingaite (kufamba chitima, nharaunda dzemakambani dzakaganhurirwa, maLab asina network), kuve neAI yemuno inopindura nekukurumidza uye isina miganho yekushandisa ibhenefiti chaiyo.
LM Studio: Dashboard yeAI yako pane macOS
LM Studio inoita sea centralized workshop yekutsvaga, kurodha uye kumhanya Mienzaniso yemitauro yemuno ine chimiro chakareruka. Pasina kuve mugadziri wepurogiramu, unogona kugadzirisa toni, tekinoroji nhanho, kana kureba kweiyo modhi, kana kusiya zvese sekumisikidza uye wotanga kuishandisa mumaminitsi.
Maitiro acho anoshanda kwazvo: kubva pakutsvaga tebhu unowana mamodheru, dhawunirodha nekudzvanya kumwe chete, uye woaisa mundangariro kuti utaure nekukurumidza. Iyo zvakare inobvumidza iwe kugadzirisa akakosha ma parameter akadai takura paGPU (VRAM yekushandisa), CPU threads, kureba kwechirevo (muDeepSeek R1 distill inosvika nhamba senge 131.072 tokens) uye sarudzo dzekurangarira senge. Chengeta Muenzaniso Mundangariro o Edza mmap()Kana iwe usingade kuomesa zvinhu, ingo shandisa zvigadziriso zvekutanga uye tinya "Load Model."
Isa uye shandisa DeepSeek R1 muLM Studio nhanho nhanho
Nzira iri nyore yekutanga paMac ndeyekurodha LM Studio, tsvaga iyo chaiyo modhi, uye kuiisa. Iyo app inokutungamira iwe mukuita uye, Kana midziyo yako isingakodzeri imwe modhi, inotoratidza chenjedzo senge "Ingangove yakakurisa pamuchina uyu".
Danho 1: Tsvaga modhi. Vhura LM Studio uye enda ku "Discover" kana "Model Search." Nyora "zvakadzama" mubhari yekutsvaga uye tsvaga "DeepSeek R1 Distill (Qwen 7B)”. Mupaneru yekurudyi uchaona chiyero chakakura (semuenzaniso, 4,68 GB) nemaitiro ayo. inoshanda chaizvo pakufunga uye inokwana zvakanaka mumakomputa ane ndangariro shoma.
Danho 2: dhawunirodha. Dzvanya bhatani regirini "Download" (uchaona saizi, semuenzaniso, 4,68 GB). Iwindo reparutivi rinoratidza kufambira mberi, kumhanya, uye nguva inofungidzirwa. Kana yapera, modhi yacho ichaonekwa mune yako runyorwa rwekudhaunirodha iripo. Inosimbisa kuti "DeepSeek R1 Distill Qwen 7B" inooneka asati aenderera mberi.
Danho rechitatu: Rongedza uye takura. Usati wadzvanya "Road Model", unogona kugadzirisa: kureba kwechirevo, VRAM yekushandisa, nhamba yeCPU tambo, uye kana kuchengeta modhi mundangariro. Kuti utange, iyo default setting ichashanda. Nekudaro, kana yako Mac Mini ine yakakwana yakabatana ndangariro, ipfungwa yakanaka kugonesa "Chengeta Muenzaniso Mundangariro” kuitira kukurumidza kuzadza.
Danho rechina: Chat. Vhura iyo "Chat" maonero, sarudza modhi yakarodha, uye nyora chimwe chinhu senge, "Mhoro, uri modhi ipi uye ndiani akakudzidzisa?" Kana vakapindura nehunhu hwavo uye kugona kwavo, unoziva zvese zviri kushanda. Kana ukaedza modhi hombe (semuenzaniso, inoda zvikuru “DeepSeek-V3-4bit”) uye inoratidzika mutsvuku, “Pamwe wakakurisa pamushini uyu”, iwe unofanirwa kusarudza imwe yakareruka kana kuwedzera RAM/VRAM pamakomputa anoenderana.
Pakarepo uye kushanda, unogona kuishandisa. Hapana internet inobatanidzaChiyero chinobatsira chekutarisa izvi kudzima Wi-Fi muSystem Settings uye kuvhura Activity Monitor kuona mashandisirwo eGPU paunenge uchitaura nemuenzaniso; kana girafu ichifamba, basa rese rinoitwa paMac yako.
Neimwe nzira: Isa DeepSeek naOllama pane macOS
Kana iwe uchida kureruka kweiyo terminal, Ollama inonyanyozivikanwa yemuno LLM modhi server. Iwe unodhawunirodha app yayo yeMacOS, iise senge chero imwe, uye unogona kukumbira mamodheru nemurairo wakapusa.
Kuti utange, isa Ollama kubva kune yayo yepamutemo webhusaiti uye imhanye. MuTerminal, murairo wakajairika weiyo diki vhezheni ndeiyi: ollama run deepseek-r1:7b. Kana iwe uine yakawanda ndangariro (semuenzaniso 32 GB kana kupfuura) unogona kuedza yakakwira akasiyana (13B kana kuti, kana ukashinga, 67B), kunyangwe paMac Mini chiitiko chacho chinowanzo kugadzikana ne7B kana 8B.
Avo vanoda imwe yekuona interface vanogona kuibatanidza pamusoro peOllama. Dzimwe sarudzo dziripo Chatbox AI (iwe unongedza mupi ku "Ollama API" uye sarudza "DeepSeek R1 7B") kana kuwedzera Peji Rubatsiro - Iyo Webhu UI yemunharaunda AI Models, iyo inopa ChatGPT-chimiro chepaneru mubrowser asi kuvimba neAI yako yemuno.
Kuti uone kuti zvese ndezvenzvimbo zvechokwadi, dzima Wi-Fi, mhanyisa mubvunzo, uye tarisa zvakare Activity Monitor muGPU tab. Iwe uchaona iyo sisitimu uchishandisa iyo yakasanganiswa magiraidhi kana Apple Silicon's yakabatana ndangariro. zvichisimbisa kuti hakuna traffic kune gore.
R1 Distillate kana V3? Saizi, Hardware, uye modhi sarudzo
DeepSeek R1 muzvinyorwa zvakasvibiswa (zvakadai seQwen 7B kana Llama 8B-based sarudzo) ndiyo inotapira nzvimbo yemidziyo yemumbaInochengetedza kukosha kwekufunga uku ichidzikisa tsoka: isu tiri kutaura mapakeji ari pakati pe4 ne8 GB, inogoneka kwazvo kuMac Mini ine 16 GB ye RAM yakabatana.
Iyo yepamusoro-ye-iyo-range yakazara shanduro, DeepSeek-R1:671B, is data center material. Yakadzvanywa, inogona kunge yakatenderedza 120 GB (ine mabviro mumazana eGB) uye kuita kwayo chaiko kunoda. akawanda nyanzvi-giredhi GPUs nemari yakawanda yeVRAM. Kuenzanisira, kune makore demos pamanode ane masere 192GB AMD Instinct MI300X GPUs, gumi nemaviri kana mazana eCPU cores, uye terabytes eRAM. Iyi haisi iyo desktop yevatengi.
Kana uri kushandisa LM Studio, ndapota tarisa kuenderana kwayo zviziviso. Kana uri kushandisa Ollama, ndokumbira uise pamberi 7B; kana ikafamba zvakanaka uye kushandisa kwako kuchida, edza 13B. PaApple Silicon, kushanda kwesimba kunopenya, uye kunyangwe isina dGPU, Zviratidzo pasekondi yakanaka kwazvo. yekunyora, kufunga, kuronga mwenje uye kubvunzana kwehunyanzvi.
Kuita chaiko kwepasirese paMac Mini uye mamwe maMac ane Apple Silicon
Miedzo muna a Mac mini ine M4 chip uye 16GB ratidza kuti madiki/ epakati modhiyo emuno anopindura nekukurumidza. Kunyangwe pasina girafu yakatsaurirwa, iyo yakabatana memory uye accelerators Iko kugona kweSoC inobvumira kukurumidza, yakaderera-latency mameseji kugadzirwa pane zvakajairika kukurudzira.
Mukuenzanisa zvisina kurongwa newebhu sarudzo dzakadai seChatGPT (GPT-4), Claude 3.5 Sonnet, Gemini 1.5 Flash kana iyo online DeepSeek V3 pachayo, mhando dzemuno senge Llama 3.1-8B, Phi-4-14B o DeepSeek R1-14B Ivo vanokatyamadza nekumhanya kwavo kwekupindura, kunyangwe vachimhanya panguva imwe chete. Nekudaro, kana wakatarisana nemitoro inorema kana kureba kukurudzira, Gore rinoramba richiwanda mumhasuru mbishi.
Kuyera "tokens pasekondi" munharaunda kunobatsira pakuongorora kuvandudzwa kwehardware kana kusarudza kuti ungakwidziridze kune yakakura modhi saizi. NeLM Studio naOllama, zviri nyore kudzokorora kukurumidza kumwechete uye rekodha kuita kuenzanisa zvigadziriso (zvakawanda zveCPU tambo, VRAM mutsauko, kureba kwechirevo, nezvimwe).
Zvavanowana zvakanaka (uye zvavasina): kufunga, chokwadi, uye kusarura
"Kukurukurirana" bvunzo dzakadai sekuverenga mabhii zvinopa mhedzisiro inonakidza. Nezwi "Strawberry", mamwe mamodheru emuno anotadza kana kumhanyisa, nepo DeepSeek R1 distillation inogona dyara nguva yakawanda uchifunga asi zvigadzirise, uchitsanangura nhanho nhanho kuti ungawedzera sei "r".
Muchirevo cheSpanish "Imbwa yaSaint Roque haina muswe nekuti Ramón Ramírez akaiba," mamiriro acho anowedzera kuoma: akati wandei ewebhu chatbots anoita zvikanganiso pakutanga uye, mushure mekubvunzwa zvakare, vanozvigadzirisa. Munzvimbo, R1 uye mamwe mamodheru Vanogona kuvhiringika nemutauro kana nechinangwa chekutsvaga (kuvhiringidza "r" ne "e"), izvo zvinojekesa kuti zvinokurudzirwa kuvatungamirira uye, kana zvichidiwa, edza zvakare nemirayiridzo yakanyatsojeka.
Nemapikicha ekufunga kwepashure, seaya ari pamusoro pemabhora ebhiriyadhi 7-9-11-13 achiwedzera kusvika pamakumi matatu kana ukapeta 9 kuita 6, mhinduro isina kuzara kazhinji "haugone." Kunyangwe iwe ukasimbirira pane "kune hunyengeri," Vazhinji vemo havawani nzira yekugadzira, nepo mamwe masevhisi ewebhu, ari munzira, anogadzirisa.
Pamibvunzo yechokwadi (semuenzaniso, tafura yevakundi veWorld Cup nevamhanyi), Cloud masevhisi akairoverera mubhuraketi richangoburwa, nepo pachinzvimbo pakanga paine kufungidzira uye data isiriyo (vakagadzira mafinalists, mazita akagoverwa zvisizvo, nezvimwewo). Pano kurudziro ndeyekusiyanisa uye, kana iwe uchida nhoroondo yechokwadi, vimba nemanyuko akasimbiswa kana kushandisa hombe/yakagadziridzwa modhi.
Paunenge uchitaura nyaya dzinonetsa (Tiananmen, Taiwan, kushoropodzwa kwevatungamiriri), nuances inofarirwa: mamwe mawebhu mamodheru anorambidza zvirimo uye DeepSeek R1 munharaunda inogona. pindura nemasefa kana mapfupi zvichienderana nekukasira, dzimwe nguva nemameseji mune mimwe mitauro panguva yavo "yekufunga." Rutivi rwakanaka rwezvinhu nderokuti, kazhinji, vagari vomunharaunda vanodzorwa uye vanoremekedza, uye vanodzivisa mirairo ine ngozi (yakadai sekugadzira bhomba), kuramba nenyevero dzine musoro.
Kuvanzika uye Zvakaitika Zvemunharaunda: Zvaunoda Kuziva
Iyo nharo huru yekuisa DeepSeek paMac yako ndiyo Iko hakuna mapato echitatu ari kuverenga chats dzakoIwe hauvimbe ne quotas kana miganho yekushandisa, uye unogona kuramba uchishanda pasina kuvharwa. Nekudaro, kana iwe ukatsvaga mawebhusaiti ekuenzanisira kana maforamu, uchaona zviziviso zvecookie (sezviya zviri paReddit) uchikumbira mvumo yako; izvi zvinongokanganisa chikuva chavo. kwete kuurayiwa kwaunogara.
Pa-nzvimbo, zvese zviri pasi pesimba rako: unogona kuchengetedza nhaurirano, gadzirisa ma paramita, uye sarudza nguva yekuvandudza kana kuchinja mamodheru. Uyezve, zvishandiso zvakafanana AnythingLLM o Chero kupi LLM Vanopa kumwe kuyerera nemaseva emunharaunda uye, pazvinenge zvakakodzera, mawebhusaiti. zvakafanana neaya epamhepo chatbots asi pasina kutora data rako kunze.
Mazano ekuseta: Wana Zvakawanda Kunze kweMac Yako
Kana iwe ukasangana nekunonoka, dzikisa saizi yemuenzaniso kana shandisa zvakasiyana. 4-bit quantized kana iripo. MuLM Studio, kana ukasangana ne "Ingangove yakakurisa pamuchina uyu," usamanikidzire: musangano wakagadzikana une 7B/8B yakanyatsogadzirwa. inobereka zvinopfuura 13B pamuganho.
Activate "Chengeta Muenzaniso Mundangariro” kukurumidza kushandura pakati pechats nemhando “Edza mmap()” kana puratifomu ichiitsigira; paApple Silicon inobatsira nekurangarira manejimendi. Isa huwandu hwetambo dzeCPU kusvika pahafu kana matatu muzvina zvemacores ako kuti usazadze mamwe mabasa, uye ugogovera VRAM nekuchengetedza kana iwe ukashanda nemaapuro anorema akafanana (kugadziridza vhidhiyo, maIDE, nezvimwewo).
Hurefu hwechinyorwa hunoyedza (semuenzaniso, 131.072 zviratidzo), asi haudi nguva dzese. Kuideredza kunovandudza latency uye kushandiswa kwesimba. Chengetedza zvimiro zvakakura zve magwaro marefu kana kuongorora kwekodhi, uye shandisa zvirevo zvepfupiso zuva nezuva.
Kana iwe uri kuyedza nemamodhi akawanda kamwechete, dzivisa kuamhanyisa achienderana pamichina ye16GB; dzimwe nguva kana kuvhara yausingashandisi kudzorera ndangariro dzakabatana kuhurongwa. Tarisa Chiitiko Monitor: Kana iwe ukaona yakakwirira ndangariro kudzvanywa, inguva yekuburitsa modhi, kusunungura zviwanikwa, kana kutevera a. Nhungamiro yekuti kana Mac yako yakavharwa.
Inokurudzirwa workflows
Kana iwe uchida ChatGPT-chimiro chechimiro pamusoro peOllama, shandisa Chatbox AI kana Peji Rubatsiro. Gadzirisa "Ollama API," sarudza "DeepSeek R1 7B," uye ndizvozvo: unenge uine hwindo rakachena uko. Mayedzo ekukurudzira, chengetedza zvikamu, uye chinja modhi pasina mirairo yekunyora.
Mibvunzo inokurumidza
- Ndinogona kushandisa DeepSeek pasina Internet? Ehe. Dhawunirodha modhi uye, kana yangotakura, unogona kudzima Wi-Fi. Iko kuverenga kunoitwa paMac yako, uye iyo data haina kubuda.
- Ndeipi Mac yandinoda? NeApple Silicon (M-series) uye 16GB yendangariro yakabatana, 7GB/8GB inoshanda chaizvo. Unogona kuedza 8GB, asi iwe unenge uine shoma headroom.
- Yakarurama here segore? Zvinoenderana nenyaya. Inoita zvakanaka kwazvo pakufunga uye kugadzira; kune data yakakosha yechokwadi, simbisa zvinyorwa kana kushandisa mhando hombe.
- Ndinogona kuchinja mamodheru nyore nyore? Ehe. LM Studio neOllama inokutendera iwe kudhawunirodha uye chinja pakati pemhando (Llama, Phi, Qwen, nezvimwewo) kuenzanisa kuita uye maitiro.
Kukwidza DeepSeek pane yako Mac Mini inzira inobatsira yekuunza AI kune desktop yako nayo kumhanya, kuvanzika uye kutonga kwakazaraNeLM Studio une ruzivo rwakatungamirirwa, uye naOllama une kuyerera kwakareruka pane imwe terminal; ese ari maviri anokupa iwe kuwana kune anokwanisa kwazvo akadhindwa shanduro, zvinogutsa kufunga mumabasa mazhinji, uye simba rakakwana rekunyora, chirongwa, uye kuyedza, uchiziva izvozvo. zvese zvinogadziriswa zvinoitika pakombuta yako.
