{"id":29372,"date":"2026-10-23T22:24:45","date_gmt":"2026-10-23T22:24:45","guid":{"rendered":"https:\/\/www.websiteradar.net\/?p=29372"},"modified":"2026-10-23T22:24:45","modified_gmt":"2026-10-23T22:24:45","slug":"helyi-ai-hardver","status":"publish","type":"post","link":"https:\/\/www.websiteradar.net\/hu\/helyi-ai-hardver\/","title":{"rendered":"Helyi AI-hardver 2026-ban: milyen sz\u00e1m\u00edt\u00f3g\u00e9p kell az LLM-ek otthoni futtat\u00e1s\u00e1hoz"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\">Saj\u00e1t g\u00e9pen szeretn\u00e9l nyelvi modelleket futtatni \u2013 el\u0151fizet\u00e9s n\u00e9lk\u00fcl, \u00fagy, hogy az adataid nem hagyj\u00e1k el az asztalodat, \u00e9s m\u00e9g internetkapcsolat sem kell hozz\u00e1? Akkor a processzor \u00e9s a dobozon d\u00edszelg\u0151 \u201eTOPS\u201d matrica sokkal kev\u00e9sb\u00e9 sz\u00e1m\u00edt, mint k\u00e9t m\u00e1sik dolog: <strong>mennyi mem\u00f3ri\u00e1t haszn\u00e1lhat a modell<\/strong>, \u00e9s <strong>milyen gyorsan tudja a g\u00e9p olvasni ezt a mem\u00f3ri\u00e1t<\/strong>. Ez az \u00fatmutat\u00f3 elmagyar\u00e1zza, hogyan fut val\u00f3j\u00e1ban a helyi AI egy sz\u00e1m\u00edt\u00f3g\u00e9pen, mennyi mem\u00f3ri\u00e1t ig\u00e9nyelnek a n\u00e9pszer\u0171 modellek, \u00e9s hogyan teljes\u00edtenek a gyakorlatban a mai hardverek \u2013 a megl\u00e9v\u0151 PC-be tehet\u0151 videok\u00e1rty\u00e1t\u00f3l eg\u00e9szen az eg\u00e9sz csapatot kiszolg\u00e1l\u00f3 munka\u00e1llom\u00e1sig. Minden eszk\u00f6zh\u00f6z ellen\u0151rz\u00f6tt sz\u00e1mokat, forr\u00e1sokat \u00e9s egy\u00e9rtelm\u0171 \u00edt\u00e9letet adunk arr\u00f3l, kinek val\u00f3; n\u00e9h\u00e1ny fejezethez r\u00f6vid vide\u00f3 is tartozik.<\/p>\n\n\n\n<div class=\"wp-block-group dw-summary has-base-2-background-color has-background\" style=\"border-radius:16px;padding-top:24px;padding-right:28px;padding-bottom:24px;padding-left:28px\"><div class=\"wp-block-group__inner-container is-layout-constrained wp-container-core-group-is-layout-bba1aa89 wp-block-group-is-layout-constrained\">\n\n<h3 class=\"wp-block-heading\">A l\u00e9nyeg 30 m\u00e1sodpercben<\/h3>\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>A mem\u00f3ria d\u00f6nt mindenr\u0151l.<\/strong> A teljes modellnek bele kell f\u00e9rnie egy olyan gyors mem\u00f3ri\u00e1ba, amelyet a GPU el\u00e9r. 4 bites kvant\u00e1l\u00e1sn\u00e1l sz\u00e1molj nagyj\u00e1b\u00f3l <strong>milli\u00e1rd param\u00e9terenk\u00e9nt 0,6&nbsp;GB-tal<\/strong>.<\/li>\n\n\n<li><strong>A mem\u00f3ria-s\u00e1vsz\u00e9less\u00e9g hat\u00e1rozza meg, milyen gyorsan \u00edr a modell<\/strong>, a sz\u00e1m\u00edt\u00e1si teljes\u00edtm\u00e9ny pedig azt, milyen gyorsan olvassa be a hossz\u00fa promptot. Ez\u00e9rt \u00edr gyorsabban a Mac, mint a DGX Spark, a Spark viszont gyorsabban dolgozza fel a hossz\u00fa dokumentumokat.<\/li>\n\n\n<li><strong>Jelenleg a legjobb \u00e1r-\u00e9rt\u00e9k:<\/strong> Mac Studio M5 Max 128&nbsp;GB-tal (gyors kimenet), Ryzen AI Max+ 395&nbsp;128&nbsp;GB-tal (a legolcs\u00f3bb nagy mem\u00f3ria), RTX 5060 Ti 16&nbsp;GB (a legolcs\u00f3bb \u00e9rtelmes bel\u00e9p\u0151).<\/li>\n\n\n<li><strong>Az NPU-val szerelt \u201eAI PC\u201d nem helyi AI-ra val\u00f3 sz\u00e1m\u00edt\u00f3g\u00e9p.<\/strong> Sem az Ollama, sem az LM Studio nem haszn\u00e1lja az NPU-t.<\/li>\n\n<\/ul>\n\n<\/div><\/div>\n\n\n\n<p class=\"dw-disclosure has-small-font-size wp-block-paragraph\">Az eszk\u00f6z\u00f6ket nem tesztelt\u00fck saj\u00e1t magunk \u2013 a sz\u00e1mok a cikk v\u00e9g\u00e9n felsorolt forr\u00e1sokb\u00f3l sz\u00e1rmaznak, a k\u00f6z\u00f6ss\u00e9gi m\u00e9r\u00e9seket k\u00fcl\u00f6n jel\u00f6lj\u00fck. Az \u00e1rak d\u00e1tummal ell\u00e1tott amerikai lista\u00e1rak; a mem\u00f3riahi\u00e1ny miatt 2026 folyam\u00e1n sok k\u00f6z\u00fcl\u00fck emelkedett.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"podle-rozpoctu\">Mit \u00e9rdemes venni: gyors aj\u00e1nl\u00f3<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Ha nem akarod v\u00e9gigolvasni az eg\u00e9sz \u00fatmutat\u00f3t, itt a r\u00f6vid\u00edtett v\u00e1ltozat. Minden aj\u00e1nl\u00e1st lejjebb r\u00e9szletesen is bemutatunk, n\u00e9h\u00e1nyat vide\u00f3val.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><a href=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.hu_-scaled.webp\"><img loading=\"lazy\" decoding=\"async\" width=\"2560\" height=\"1963\" src=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.hu_-scaled.webp\" alt=\"Mit vegy\u00e9l helyi AI-hoz, a bel\u00e9p\u0151 szintt\u0151l a munka\u00e1llom\u00e1sig\" class=\"wp-image-29364\" srcset=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.hu_-scaled.webp 2560w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.hu_-300x230.webp 300w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.hu_-1024x785.webp 1024w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.hu_-768x589.webp 768w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.hu_-1536x1178.webp 1536w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.hu_-2048x1571.webp 2048w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/rozpocet.hu_-610x468.webp 610w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/><\/a><figcaption class=\"wp-element-caption\">Gyors aj\u00e1nl\u00f3 kateg\u00f3ri\u00e1nk\u00e9nt, a legolcs\u00f3bb \u00e9rtelmes bel\u00e9p\u0151t\u0151l a csapatmunka\u00e1llom\u00e1sig.<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table class=\"has-fixed-layout\"><thead><tr><th>Kateg\u00f3ria<\/th><th>Aj\u00e1nlat<\/th><th>Milyen modellekhez<\/th><\/tr><\/thead><tbody><tr><td>Bel\u00e9p\u0151<\/td><td><a href=\"#rtx-5060-ti-16gb\">RTX 5060 Ti 16&nbsp;GB<\/a> a megl\u00e9v\u0151 PC-be<\/td><td>8\u201320B (Qwen3 14B, gpt-oss-20b)<\/td><\/tr><tr><td>Kezd\u0151<\/td><td><a href=\"#mac-mini-m6\">Mac mini M6 32&nbsp;GB-tal<\/a> vagy <a href=\"#rtx-5070-ti-vs-5080\">RTX 5070 Ti<\/a><\/td><td>14B-ig gyorsan, Qwen3 30B-A3B<\/td><\/tr><tr><td>64&nbsp;GB<\/td><td><a href=\"#mac-mini-m5-pro\">Mac mini M5 Pro 64&nbsp;GB-tal<\/a><\/td><td>20\u201335B, csendes otthoni szerver<\/td><\/tr><tr><td>128&nbsp;GB, legjobb \u00e1r-\u00e9rt\u00e9k<\/td><td><a href=\"#ryzen-ai-max-395\">Ryzen AI Max+ 395&nbsp;128&nbsp;GB-tal<\/a><\/td><td>nagy MoE-modellek 120B-ig<\/td><\/tr><tr><td>128&nbsp;GB, leggyorsabb kimenet<\/td><td><a href=\"#mac-studio-m5-max\">Mac Studio M5 Max 128&nbsp;GB-tal<\/a><\/td><td>gpt-oss-120b 100 token\/s felett<\/td><\/tr><tr><td>CUDA-fejleszt\u0151i g\u00e9p<\/td><td><a href=\"#dgx-spark\">DGX Spark vagy m\u00e1s GB10-es g\u00e9p<\/a><\/td><td>CUDA-fejleszt\u00e9s, k\u00f3dol\u00f3 \u00fcgyn\u00f6k\u00f6k<\/td><\/tr><tr><td>Maxim\u00e1lis mem\u00f3ria<\/td><td><a href=\"#mac-studio-m5-ultra\">Mac Studio M5 Ultra 256&nbsp;GB-tal<\/a><\/td><td>100\u2013250B-s modellek otthon<\/td><\/tr><tr><td>Csapat-munka\u00e1llom\u00e1s<\/td><td><a href=\"#rtx-pro-6000\">RTX PRO 6000<\/a><\/td><td>70\u2013120B egy eg\u00e9sz csapatnak<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"proc-lokalni-ai\">Mi\u00e9rt \u00e9rdemes helyben futtatni az AI-t \u2013 \u00e9s mit ne v\u00e1rj t\u0151le<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A helyi modell a saj\u00e1t sz\u00e1m\u00edt\u00f3g\u00e9peden fut. Az adataid nem hagyj\u00e1k el az asztalodat, nem fizetsz havi el\u0151fizet\u00e9st vagy tokenalap\u00fa d\u00edjat, a modell internet n\u00e9lk\u00fcl is m\u0171k\u00f6dik, \u00e9s senki sem v\u00e1ltoztathatja meg vagy vezetheti ki egyik napr\u00f3l a m\u00e1sikra. Ez\u00e9rt j\u00f3 v\u00e1laszt\u00e1s \u00e9rz\u00e9keny dokumentumokhoz, c\u00e9ges k\u00f3don dolgoz\u00f3 k\u00f3dol\u00e1si asszisztensekhez, naponta t\u00f6bb ezer k\u00e9r\u00e9st k\u00fcld\u0151 automatiz\u00e1ci\u00f3khoz \u2013 vagy egyszer\u0171en a k\u00eds\u00e9rletez\u00e9s \u00f6r\u00f6m\u00e9\u00e9rt.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">K\u00e9t b\u00f6kken\u0151 van. Az egyik, hogy az otthoni hardverre f\u00e9r\u0151 ny\u00edlt modellek okosak, de nem \u00e9rik el a legjobb felh\u0151szolg\u00e1ltat\u00e1sok szintj\u00e9t. A m\u00e1sik, hogy a hardver nem olcs\u00f3, \u00e9s 2026-ban a mem\u00f3riahi\u00e1ny felfel\u00e9 hajtja az \u00e1rakat. A helyi futtat\u00e1s mellett els\u0151sorban az adatv\u00e9delem, a kontroll \u00e9s a f\u00fcggetlens\u00e9g sz\u00f3l \u2013 nem a sp\u00f3rol\u00e1s.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"jak-model-bezi\">Hogyan fut a modell a g\u00e9peden<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Amikor k\u00e9rd\u00e9st k\u00fcldesz a modellnek, k\u00e9t f\u00e1zis zajlik le, \u00e9s mindkett\u0151t m\u00e1s korl\u00e1tozza:<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n\n<li><strong>A prompt beolvas\u00e1sa (prefill).<\/strong> A modell egyszerre dolgozza fel a teljes bemenetet \u2013 a k\u00e9rd\u00e9st, a bem\u00e1solt dokumentumot, a projekted k\u00f3dj\u00e1t. Ez a f\u00e1zis a GPU <strong>sz\u00e1m\u00edt\u00e1si teljes\u00edtm\u00e9ny\u00e9n<\/strong> m\u00falik, \u00e9s ez hat\u00e1rozza meg, mennyit v\u00e1rsz a v\u00e1lasz els\u0151 szav\u00e1ra. Egy r\u00f6vid k\u00e9rd\u00e9sn\u00e9l a m\u00e1sodperc t\u00f6red\u00e9k\u00e9ig tart, egy t\u00f6bb t\u00edzezer szavas dokumentumn\u00e1l ak\u00e1r percekig is.<\/li>\n\n\n<li><strong>A v\u00e1lasz \u00edr\u00e1sa (decode).<\/strong> A modell szavank\u00e9nt (pontosabban tokenenk\u00e9nt) gener\u00e1lja a sz\u00f6veget. Minden tokenhez be kell olvasnia a mem\u00f3ri\u00e1b\u00f3l a teljes modellt \u2013 a szak\u00e9rt\u0151kever\u00e9ket (mixture of experts) haszn\u00e1l\u00f3 modellekn\u00e9l csak az akt\u00edv r\u00e9sz\u00e9t. Ez a f\u00e1zis a <strong>mem\u00f3ria-s\u00e1vsz\u00e9less\u00e9gen<\/strong> m\u00falik: h\u00e1ny gigab\u00e1jtot tud a chip m\u00e1sodpercenk\u00e9nt beolvasni.<\/li>\n\n<\/ol>\n\n\n\n<figure class=\"wp-block-image size-full\"><a href=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.hu_-scaled.webp\"><img loading=\"lazy\" decoding=\"async\" width=\"2560\" height=\"1341\" src=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.hu_-scaled.webp\" alt=\"A prompt beolvas\u00e1sa \u00e9s a v\u00e1lasz \u00edr\u00e1sa: k\u00e9t k\u00fcl\u00f6nb\u00f6z\u0151 sz\u0171k keresztmetszet\" class=\"wp-image-29365\" srcset=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.hu_-scaled.webp 2560w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.hu_-300x157.webp 300w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.hu_-1024x536.webp 1024w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.hu_-768x402.webp 768w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.hu_-1536x804.webp 1536w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.hu_-2048x1073.webp 2048w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/jak-model-bezi.hu_-610x319.webp 610w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/><\/a><figcaption class=\"wp-element-caption\">P\u00e9lda: gpt-oss-120b a llama.cpp-ben. A prompt beolvas\u00e1s\u00e1ban az NVIDIA, az \u00edr\u00e1sban az Apple vezet.<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">A gyakorlati k\u00fcl\u00f6nbs\u00e9g: <strong>cseveg\u00e9sn\u00e9l<\/strong> a gyors kimenetet \u00e9rzed (\u00e9s itt brill\u00edrozik a Mac). <strong>K\u00f3dol\u00f3 \u00fcgyn\u00f6k\u00f6kn\u00e9l \u00e9s dokumentumokkal v\u00e9gzett munk\u00e1n\u00e1l<\/strong>, amelyek folyamatosan hatalmas kontextust k\u00fcldenek a modellnek, a gyors promptbeolvas\u00e1s ugyanolyan fontos (\u00e9s itt az NVIDIA vezet). A sebess\u00e9get token m\u00e1sodpercenk\u00e9nt (token\/s) m\u00e9rj\u00fck \u2013 egy token nagyj\u00e1b\u00f3l h\u00e1romnegyed sz\u00f3. K\u00f6r\u00fclbel\u00fcl 15 token\/s felel meg az olvas\u00e1si sebess\u00e9gnek; 50 felett a modell f\u00fcrg\u00e9nek \u00e9rz\u0151dik.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"kolik-pameti\">Mennyi mem\u00f3ria kell<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A teljes modellnek olyan gyors mem\u00f3ri\u00e1ban kell \u00fclnie, amelyhez a GPU hozz\u00e1f\u00e9r: vagy a videok\u00e1rtya VRAM-j\u00e1ban, vagy a Macekben, a DGX Sparkban \u00e9s a Ryzen AI Maxban tal\u00e1lhat\u00f3 k\u00f6z\u00f6s (egyes\u00edtett, unified) mem\u00f3ri\u00e1ban. Ha nem f\u00e9r bele, egy r\u00e9sze \u00e1tker\u00fcl a norm\u00e1l rendszermem\u00f3ri\u00e1ba, \u00e9s a sebess\u00e9g \u00f6sszeomlik. Egy RTX 5060 Ti-on v\u00e9gzett teszten m\u00e1r az is, hogy a Qwen 14B r\u00e9tegeinek csak a hatod\u00e1t tett\u00e9k \u00e1t a GPU-r\u00f3l, <strong>43-r\u00f3l 12,5 token\/s-ra<\/strong> vetette vissza a kimeneti sebess\u00e9get; puszt\u00e1n a processzoron 2,9 token\/s j\u00f6tt ki.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><a href=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.hu_-scaled.webp\"><img loading=\"lazy\" decoding=\"async\" width=\"2560\" height=\"1890\" src=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.hu_-scaled.webp\" alt=\"N\u00e9pszer\u0171 modellek m\u00e9rete 4 biten az eszk\u00f6z\u00f6k mem\u00f3ri\u00e1j\u00e1hoz k\u00e9pest\" class=\"wp-image-29366\" srcset=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.hu_-scaled.webp 2560w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.hu_-300x221.webp 300w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.hu_-1024x756.webp 1024w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.hu_-768x567.webp 768w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.hu_-1536x1134.webp 1536w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.hu_-2048x1512.webp 2048w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/velikost-modelu.hu_-610x450.webp 610w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/><\/a><figcaption class=\"wp-element-caption\">GGUF-f\u00e1jlm\u00e9retek a Hugging Face-r\u0151l. Ehhez j\u00f6n m\u00e9g a kontextus mem\u00f3riaig\u00e9nye.<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table class=\"has-fixed-layout\"><thead><tr><th>Modell<\/th><th>T\u00edpus<\/th><th>M\u00e9ret (4 bit)<\/th><th>Hov\u00e1 f\u00e9r el<\/th><\/tr><\/thead><tbody><tr><td>Llama 3.1 8B \/ Qwen3 8B<\/td><td>s\u0171r\u0171<\/td><td>4,9\u20135,0&nbsp;GB<\/td><td>b\u00e1rmibe, aminek 8&nbsp;GB-ja van<\/td><\/tr><tr><td>Qwen3 14B<\/td><td>s\u0171r\u0171<\/td><td>9,0&nbsp;GB<\/td><td>12\u201316&nbsp;GB<\/td><\/tr><tr><td>gpt-oss-20b<\/td><td>MoE<\/td><td>12,1&nbsp;GB (MXFP4)<\/td><td>16&nbsp;GB<\/td><\/tr><tr><td>Gemma 3 27B<\/td><td>s\u0171r\u0171<\/td><td>16,6&nbsp;GB<\/td><td>24&nbsp;GB<\/td><\/tr><tr><td>Qwen3 30B-A3B<\/td><td>MoE, 3B akt\u00edv<\/td><td>18,6&nbsp;GB<\/td><td>24&nbsp;GB<\/td><\/tr><tr><td>Qwen3 32B<\/td><td>s\u0171r\u0171<\/td><td>19,8&nbsp;GB<\/td><td>24\u201332&nbsp;GB<\/td><\/tr><tr><td>Llama 3.3 70B<\/td><td>s\u0171r\u0171<\/td><td>42,5&nbsp;GB<\/td><td>48&nbsp;GB vagy t\u00f6bb<\/td><\/tr><tr><td>gpt-oss-120b<\/td><td>MoE<\/td><td>63,4&nbsp;GB (MXFP4)<\/td><td>96\u2013128&nbsp;GB<\/td><\/tr><tr><td>GLM-4.5-Air 106B<\/td><td>MoE<\/td><td>73,0&nbsp;GB<\/td><td>96&nbsp;GB vagy t\u00f6bb<\/td><\/tr><tr><td>Qwen3.5&nbsp;122B-A10B<\/td><td>MoE<\/td><td>76,5&nbsp;GB<\/td><td>96\u2013128&nbsp;GB<\/td><\/tr><tr><td>Qwen3&nbsp;235B-A22B<\/td><td>MoE<\/td><td>142,2&nbsp;GB<\/td><td>192\u2013256&nbsp;GB<\/td><\/tr><tr><td>DeepSeek R1&nbsp;671B<\/td><td>MoE<\/td><td>404,9&nbsp;GB<\/td><td>512&nbsp;GB<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"a-kontextusnak-is-kell-mem\u00f3ria\">A kontextusnak is kell mem\u00f3ria<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">A modellen fel\u00fcl mem\u00f3ria kell a kontextusnak (a KV-gyors\u00edt\u00f3t\u00e1rnak, angolul KV cache) is \u2013 mindannak, amire a modell \u201eeml\u00e9kszik\u201d a besz\u00e9lget\u00e9sb\u0151l vagy a dokumentumb\u00f3l. A Llama 3.3 70B eset\u00e9ben ez 32&nbsp;000 tokenes kontextusn\u00e1l nagyj\u00e1b\u00f3l 10&nbsp;GB t\u00f6bbletet jelent, 128&nbsp;000 tokenn\u00e9l pedig k\u00f6r\u00fclbel\u00fcl 40&nbsp;GB-ot. A gpt-oss modellek a cs\u00fasz\u00f3ablaknak k\u00f6sz\u00f6nhet\u0151en sokkal takar\u00e9kosabbak: a gpt-oss-120b-nek m\u00e9g 128&nbsp;000 tokenhez is legfeljebb k\u00f6r\u00fclbel\u00fcl 9&nbsp;GB kell. A kontextus n\u00f6veked\u00e9s\u00e9vel a kimeneti sebess\u00e9g is cs\u00f6kken \u2013 a gpt-oss-120b egy M5 Maxon r\u00f6vid kontextussal k\u00f6r\u00fclbel\u00fcl 80 token\/s-mal \u00edr, 128&nbsp;000 tokenn\u00e9l viszont m\u00e1r csak 14-gyel.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"nem-minden-mem\u00f3ria-jut-a-gpu-nak\">Nem minden mem\u00f3ria jut a GPU-nak<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>A macOS<\/strong> alap\u00e9rtelmez\u00e9s szerint a RAM nagyj\u00e1b\u00f3l 75%-\u00e1t engedi \u00e1t a GPU-nak (az Apple nem k\u00f6zli a pontos \u00e9rt\u00e9ket). Egy 64&nbsp;GB-os Mac \u00edgy k\u00f6r\u00fclbel\u00fcl 48&nbsp;GB-ot k\u00edn\u00e1l a modellnek, egy 128&nbsp;GB-os k\u00f6r\u00fclbel\u00fcl 96&nbsp;GB-ot. A korl\u00e1tot a <code>sudo sysctl iogpu.wired_limit_mb=\u2026<\/code> paranccsal megemelheted, de \u00fajraind\u00edt\u00e1s ut\u00e1n vissza\u00e1ll.<\/li>\n\n\n<li><strong>A Windows a Ryzen AI Maxon<\/strong> a BIOS-ban r\u00f6gz\u00edtett ar\u00e1nyban osztja fel a mem\u00f3ri\u00e1t \u2013 a GPU 128&nbsp;GB-b\u00f3l legfeljebb 96&nbsp;GB-ot kap. Linux alatt ez k\u00f6r\u00fclbel\u00fcl 124&nbsp;GB-ig mehet.<\/li>\n\n\n<li><strong>Az NVIDIA RTX Spark notebookok<\/strong> Windows alatt alapb\u00f3l nagyj\u00e1b\u00f3l 50:50 ar\u00e1nyban osztj\u00e1k meg a mem\u00f3ri\u00e1t, k\u00e9zi be\u00e1ll\u00edt\u00e1ssal 128&nbsp;GB-b\u00f3l legfeljebb 96&nbsp;GB jut a GPU-nak. A 32&nbsp;GB-os v\u00e1ltozat csak 16&nbsp;GB-ot ad a GPU-nak.<\/li>\n\n\n<li><strong>A DGX Spark<\/strong> Linux alatt dinamikusan osztja ki a mem\u00f3ri\u00e1t, \u00edgy a modell a 128&nbsp;GB nagy r\u00e9sz\u00e9t haszn\u00e1lhatja.<\/li>\n\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"kvantizace-moe\">A kvant\u00e1l\u00e1s \u00e9s a MoE egyszer\u0171en<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>A kvant\u00e1l\u00e1s<\/strong> a modellek t\u00f6m\u00f6r\u00edt\u00e9se: param\u00e9terenk\u00e9nt 16 bit helyett 8 vagy 4 bitet haszn\u00e1lsz. 4 biten (Q4, Q4_K_M vagy MXFP4 jel\u00f6l\u00e9ssel) a modell az eredeti m\u00e9ret\u00e9nek nagyj\u00e1b\u00f3l a negyed\u00e9t foglalja el, a min\u0151s\u00e9g pedig csak kicsit romlik \u2013 ez otthoni haszn\u00e1latra \u00e9sszer\u0171 alapbe\u00e1ll\u00edt\u00e1s. A 8 bit (Q8) szinte vesztes\u00e9gmentes, de k\u00e9tszer akkora.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>A szak\u00e9rt\u0151kever\u00e9k (mixture of experts, MoE)<\/strong> olyan architekt\u00fara, amelyben a modellnek rengeteg param\u00e9tere van, de minden sz\u00f3hoz csak egy kis szelet\u00fcket haszn\u00e1lja. A gpt-oss-120b 117 milli\u00e1rd param\u00e9teres, de l\u00e9p\u00e9senk\u00e9nt csak k\u00f6r\u00fclbel\u00fcl \u00f6tmilli\u00e1rdot aktiv\u00e1l; a Qwen3 30B-A3B mind\u00f6ssze h\u00e1rommilli\u00e1rdot. Egy ilyen modellnek sok mem\u00f3ria kell, de egy j\u00f3val kisebb modell sebess\u00e9g\u00e9vel \u00edr. Ez\u00e9rt a MoE-modellek ide\u00e1lisak a nagy, de lassabb k\u00f6z\u00f6s mem\u00f3ri\u00e1j\u00fa g\u00e9pekhez \u2013 DGX Spark, Ryzen AI Max vagy Mac. A s\u0171r\u0171 (dense) modellek (Llama 70B, Qwen3 32B) viszont minden sz\u00f3hoz a teljes modellt beolvass\u00e1k, \u00e9s lass\u00fa mem\u00f3ri\u00e1n csak v\u00e1nszorognak.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"propustnost\">Mem\u00f3ria-s\u00e1vsz\u00e9less\u00e9g: mi\u00e9rt \u00edr gyorsabban a Mac, mint a Spark<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A kimeneti sebess\u00e9g el\u00e9g j\u00f3l megbecs\u00fclhet\u0151: a mem\u00f3ria-s\u00e1vsz\u00e9less\u00e9get el kell osztani azzal, mennyi adatot kell a modellnek tokenenk\u00e9nt beolvasnia. Egy 4 bites, s\u0171r\u0171 70B-s modell (~42&nbsp;GB) ez\u00e9rt egy 273&nbsp;GB\/s-os DGX Sparkon nem \u00edrhat gyorsabban nagyj\u00e1b\u00f3l 6\u20137 token\/s-n\u00e1l \u2013 a m\u00e9r\u00e9sek k\u00f6r\u00fclbel\u00fcl 4,4-et mutatnak.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><a href=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.hu_-scaled.webp\"><img loading=\"lazy\" decoding=\"async\" width=\"2560\" height=\"2074\" src=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.hu_-scaled.webp\" alt=\"Helyi AI-hardverek mem\u00f3ria-s\u00e1vsz\u00e9less\u00e9ge\" class=\"wp-image-29367\" srcset=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.hu_-scaled.webp 2560w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.hu_-300x243.webp 300w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.hu_-1024x829.webp 1024w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.hu_-768x622.webp 768w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.hu_-1536x1244.webp 1536w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.hu_-2048x1659.webp 2048w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/propustnost.hu_-610x494.webp 610w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/><\/a><figcaption class=\"wp-element-caption\">A gy\u00e1rt\u00f3k \u00e1ltal megadott mem\u00f3ria-s\u00e1vsz\u00e9less\u00e9g. Minden n\u00e9v alatt a modell \u00e1ltal haszn\u00e1lhat\u00f3 mem\u00f3riakapacit\u00e1s l\u00e1that\u00f3.<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-image size-full\"><a href=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.hu_-scaled.webp\"><img loading=\"lazy\" decoding=\"async\" width=\"2560\" height=\"2016\" src=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.hu_-scaled.webp\" alt=\"Helyi AI-hardvert\u00e9rk\u00e9p: mem\u00f3riakapacit\u00e1s \u00e9s sebess\u00e9g\" class=\"wp-image-29368\" srcset=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.hu_-scaled.webp 2560w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.hu_-300x236.webp 300w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.hu_-1024x806.webp 1024w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.hu_-768x605.webp 768w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.hu_-1536x1210.webp 1536w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.hu_-2048x1613.webp 2048w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/mapa-hardwaru.hu_-610x480.webp 610w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/><\/a><figcaption class=\"wp-element-caption\">A jobbra l\u00e9v\u0151 eszk\u00f6z\u00f6kbe nagyobb modellek f\u00e9rnek, a feljebb l\u00e9v\u0151k gyorsabban \u00edrnak.<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">A t\u00e9rk\u00e9pen j\u00f3l l\u00e1tszik, hogy t\u00f6k\u00e9letes eszk\u00f6z nincs. Az NVIDIA videok\u00e1rty\u00e1inak a leggyorsabb a mem\u00f3ri\u00e1juk, de kicsi a kapacit\u00e1suk. A DGX Spark, a Ryzen AI Max \u00e9s az RTX Spark b\u0151ven k\u00edn\u00e1l kapacit\u00e1st, de lassabb mem\u00f3ri\u00e1val. Az M5 Maxos \u2013 \u00e9s f\u0151leg az M5 Ultr\u00e1s \u2013 Mac Studio mindkett\u0151t \u00f6tv\u00f6zi, cser\u00e9be nincs CUDA, \u00e9s a promptot lassabban olvassa be, mint az NVIDIA.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"srovnani-gpt-oss\">Ugyanaz a modell, k\u00fcl\u00f6nb\u00f6z\u0151 hardver<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A legtisztess\u00e9gesebb \u00f6sszehasonl\u00edt\u00e1s ugyanazt a modellt ugyanabban a programban futtatja. Az OpenAI gpt-oss-120b modellj\u00e9t (63&nbsp;GB) \u00e9s a llama.cpp-vel \u2013 a helyi modellek futtat\u00e1s\u00e1ra legelterjedtebb programmal \u2013 k\u00e9sz\u00fclt m\u00e9r\u00e9seket v\u00e1lasztottuk.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><a href=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.hu_-scaled.webp\"><img loading=\"lazy\" decoding=\"async\" width=\"2560\" height=\"1754\" src=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.hu_-scaled.webp\" alt=\"A gpt-oss-120b sebess\u00e9ge k\u00fcl\u00f6nb\u00f6z\u0151 hardvereken\" class=\"wp-image-29369\" srcset=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.hu_-scaled.webp 2560w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.hu_-300x206.webp 300w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.hu_-1024x701.webp 1024w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.hu_-768x526.webp 768w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.hu_-1536x1052.webp 1536w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.hu_-2048x1403.webp 2048w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/gpt-oss-120b.hu_-610x418.webp 610w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/><\/a><figcaption class=\"wp-element-caption\">Kimenet \u00e9s promptbeolvas\u00e1s a llama.cpp-ben. Az M5 Max \u00e9s a Ryzen AI Max adatai k\u00f6z\u00f6ss\u00e9gi m\u00e9r\u00e9sek.<\/figcaption><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"srovnani\">Az \u00f6sszes eszk\u00f6z \u00f6sszehasonl\u00edtva<\/h2>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table class=\"has-fixed-layout\"><thead><tr><th>Eszk\u00f6z<\/th><th>Mem\u00f3ria AI-ra<\/th><th>S\u00e1vsz\u00e9less\u00e9g<\/th><th>gpt-oss-120b, kimenet<\/th><\/tr><\/thead><tbody><tr><td><a href=\"#mac-studio-m5-max\">Mac Studio M5 Max<\/a><\/td><td>~96&nbsp;GB a 128&nbsp;GB-b\u00f3l<\/td><td>614&nbsp;GB\/s<\/td><td>105 t\/s\u00b9<\/td><\/tr><tr><td><a href=\"#dgx-spark\">NVIDIA DGX Spark<\/a><\/td><td>128&nbsp;GB<\/td><td>273&nbsp;GB\/s<\/td><td>58,7&nbsp;t\/s<\/td><\/tr><tr><td><a href=\"#rtx-spark-notebooky\">RTX Spark notebookok<\/a><\/td><td>legfeljebb 96&nbsp;GB a 128&nbsp;GB-b\u00f3l<\/td><td>legfeljebb 300&nbsp;GB\/s<\/td><td>\u2013<\/td><\/tr><tr><td><a href=\"#ryzen-ai-max-395\">Ryzen AI Max+ 395<\/a><\/td><td>96&nbsp;GB (Windows) \/ ~124&nbsp;GB (Linux)<\/td><td>256&nbsp;GB\/s<\/td><td>56,6&nbsp;t\/s<\/td><\/tr><tr><td><a href=\"#mac-studio-m5-ultra\">Mac Studio M5 Ultra<\/a><\/td><td>~192&nbsp;GB a 256&nbsp;GB-b\u00f3l<\/td><td>1228&nbsp;GB\/s<\/td><td>\u2013<\/td><\/tr><tr><td><a href=\"#mac-mini-m5-pro\">Mac mini M5 Pro<\/a><\/td><td>~48&nbsp;GB a 64&nbsp;GB-b\u00f3l<\/td><td>307&nbsp;GB\/s<\/td><td>nem f\u00e9r el<\/td><\/tr><tr><td><a href=\"#mac-mini-m6\">Mac mini M6<\/a><\/td><td>~24&nbsp;GB a 32&nbsp;GB-b\u00f3l<\/td><td>170&nbsp;GB\/s<\/td><td>nem f\u00e9r el<\/td><\/tr><tr><td><a href=\"#macbook-pro-m5-max\">MacBook Pro M5 Max<\/a><\/td><td>~96&nbsp;GB a 128&nbsp;GB-b\u00f3l<\/td><td>614&nbsp;GB\/s<\/td><td>105 t\/s\u00b9<\/td><\/tr><tr><td><a href=\"#rtx-5090\">RTX 5090<\/a><\/td><td>32&nbsp;GB<\/td><td>1792&nbsp;GB\/s<\/td><td>~30 t\/s\u00b2<\/td><\/tr><tr><td><a href=\"#rtx-5060-ti-16gb\">RTX 5060 Ti 16&nbsp;GB<\/a><\/td><td>16&nbsp;GB<\/td><td>448&nbsp;GB\/s<\/td><td>nem f\u00e9r el<\/td><\/tr><tr><td><a href=\"#rtx-5070-ti-vs-5080\">RTX 5070 Ti \/ 5080<\/a><\/td><td>16&nbsp;GB<\/td><td>896 \/ 960&nbsp;GB\/s<\/td><td>nem f\u00e9r el<\/td><\/tr><tr><td><a href=\"#rtx-pro-6000\">RTX PRO 6000 Max-Q<\/a><\/td><td>96&nbsp;GB<\/td><td>1792&nbsp;GB\/s<\/td><td>170,6&nbsp;t\/s<\/td><\/tr><tr><td><a href=\"#gb10-klony\">GB10 az ASUS-t\u00f3l, MSI-t\u00f3l, Lenov\u00f3t\u00f3l\u2026<\/a><\/td><td>128&nbsp;GB<\/td><td>273&nbsp;GB\/s<\/td><td>~ mint a Spark<\/td><\/tr><tr><td><a href=\"#radeon-ai-pro-r9700\">Radeon AI PRO R9700<\/a><\/td><td>32&nbsp;GB<\/td><td>~640&nbsp;GB\/s<\/td><td>nem f\u00e9r el<\/td><\/tr><tr><td><a href=\"#ai-pc-copilot\">Copilot+ notebookok<\/a><\/td><td>~8&nbsp;GB a 16&nbsp;GB-b\u00f3l<\/td><td>~135&nbsp;GB\/s<\/td><td>nem f\u00e9r el<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u00b9 K\u00f6z\u00f6ss\u00e9gi m\u00e9r\u00e9s egy 128&nbsp;GB-os MacBook Pro M5 Maxon, llama.cpp. \u00b2 A modell egy r\u00e9sze a rendszermem\u00f3ri\u00e1ba kiszervezve.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"mac-studio-m5-max\">1. Mac Studio M5 Max \u2013 a leggyorsabb kimenet a p\u00e9nz\u00e9rt<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Az M5 Max chipes Mac Studio ma a leg\u00e9sszer\u0171bb v\u00e1laszt\u00e1s mindenkinek, aki a lehet\u0151 leggyorsabb helyi cseveg\u00e9st vagy \u00fcgyn\u00f6k\u00f6t szeretn\u00e9, \u00e9s nincs sz\u00fcks\u00e9ge CUDA-ra. A 614&nbsp;GB\/s-os mem\u00f3ria-s\u00e1vsz\u00e9less\u00e9gnek k\u00f6sz\u00f6nhet\u0151en a 128&nbsp;GB-os v\u00e1ltozat 100 token\/s felett futtatja a gpt-oss-120b-t \u2013 majdnem k\u00e9tszer olyan gyorsan, mint egy DGX Spark. Az egyes GPU-magokba \u00e9p\u00edtett \u00faj neur\u00e1lis gyors\u00edt\u00f3k a promptbeolvas\u00e1st is felgyors\u00edtott\u00e1k, de csak naprak\u00e9sz szoftverekben (l\u00e1sd a <a href=\"#software\">szoftverekr\u0151l sz\u00f3l\u00f3 r\u00e9szt<\/a>).<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Mem\u00f3ria:<\/strong> 36, 48, 64 vagy 128&nbsp;GB \u00b7 <strong>s\u00e1vsz\u00e9less\u00e9g<\/strong> 460&nbsp;GB\/s (36&nbsp;GB) \/ 614&nbsp;GB\/s<\/li>\n\n\n<li><strong>Fogyaszt\u00e1s:<\/strong> 7&nbsp;W \u00fcresj\u00e1ratban, max. 200&nbsp;W<\/li>\n\n\n<li><strong>Csatlakoz\u00f3k:<\/strong> 4\u00d7 Thunderbolt 5, 10GbE<\/li>\n\n\n<li><strong>Figyelem:<\/strong> a 36&nbsp;GB-os v\u00e1ltozatban gyeng\u00e9bb, 32 magos GPU van, \u00e9s a modellnek csak k\u00f6r\u00fclbel\u00fcl 27&nbsp;GB marad. 48&nbsp;GB-t\u00f3l felfel\u00e9 a teljes chipet kapod.<\/li>\n\n\n<li><strong>Amerikai lista\u00e1r:<\/strong> 2499 doll\u00e1rt\u00f3l (alapkonfigur\u00e1ci\u00f3, 2026. augusztus)<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Kinek val\u00f3:<\/strong> cseveg\u00e9s \u00e9s \u00fcgyn\u00f6k\u00f6k MoE-modellekkel, csendes otthoni AI-szerver, illetve azoknak, akik vide\u00f3t is v\u00e1gnak. <strong>Kinek nem:<\/strong> akinek CUDA, vLLM vagy tan\u00edt\u00e1s kell.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"dgx-spark\">2. NVIDIA DGX Spark \u2013 fejleszt\u0151i munka\u00e1llom\u00e1s, nem cseveg\u0151g\u00e9p<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">GB10 chip 128&nbsp;GB k\u00f6z\u00f6s mem\u00f3ri\u00e1val \u00e9s az NVIDIA teljes szoftverk\u00e9szlet\u00e9vel (CUDA, vLLM, TensorRT-LLM) egy 15 \u00d7 15 cm-es dobozban. B\u00e1rmely m\u00e1s kis g\u00e9pn\u00e9l gyorsabban olvassa be a promptot \u2013 a gpt-oss-120b-n 2444 token\/s-mal \u2013, de 273&nbsp;GB\/s-os s\u00e1vsz\u00e9less\u00e9ggel lassabban \u00edr, mint egy Mac: 58,7 token\/s-mal, a s\u0171r\u0171 Llama 70B-n pedig csak 4,4-del. Modellfuttat\u00e1s k\u00f6zben 60\u201390&nbsp;W-ot fogyaszt, \u00e9s k\u00e9t egys\u00e9g QSFP-k\u00e1bellel \u00f6sszekapcsolhat\u00f3, nagyj\u00e1b\u00f3l 400 milli\u00e1rd param\u00e9teres modellekig.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Mem\u00f3ria:<\/strong> 128&nbsp;GB LPDDR5x \u00b7 <strong>s\u00e1vsz\u00e9less\u00e9g<\/strong> 273&nbsp;GB\/s<\/li>\n\n\n<li><strong>GPU:<\/strong> Blackwell, 6144 CUDA-mag \u00b7 <strong>AI-teljes\u00edtm\u00e9ny<\/strong> 1 PFLOP FP4 (ritk\u00edt\u00e1ssal; a gyakorlatban nagyj\u00e1b\u00f3l a fele)<\/li>\n\n\n<li><strong>OS:<\/strong> DGX OS (Ubuntu Arm architekt\u00far\u00e1n)<\/li>\n\n\n<li><strong>Amerikai lista\u00e1r:<\/strong> 3999 doll\u00e1r a megjelen\u00e9skor (2025. okt\u00f3ber), 2026 febru\u00e1rj\u00e1t\u00f3l 4699 doll\u00e1r; a ServeTheHome 2026 okt\u00f3ber\u00e9ben arr\u00f3l sz\u00e1molt be, hogy az \u00e1r egy \u201e6950 doll\u00e1r k\u00f6r\u00fcli\u201d s\u00e1vba l\u00e9pett<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Kinek val\u00f3:<\/strong> CUDA-fejleszt\u00e9s, k\u00f3dol\u00f3 \u00fcgyn\u00f6k\u00f6k nagy kontextussal, finomhangol\u00e1s 70B-ig. <strong>Kinek nem:<\/strong> gyors cseveg\u00e9s s\u0171r\u0171 70B-s modellekkel, Windows \u00e9s j\u00e1t\u00e9k, kezd\u0151k.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"rtx-spark-notebooky\">3. NVIDIA RTX Spark notebookok \u2013 128&nbsp;GB a h\u00e1tizs\u00e1kban<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Az N1X chip Windows-notebookokhoz igaz\u00edtott GB10, gyorsabb mem\u00f3ri\u00e1val (legfeljebb 300&nbsp;GB\/s) \u00e9s 45\u201380&nbsp;W-os fogyaszt\u00e1si kerettel. A The Register tesztj\u00e9ben egy notebook 67,6 token\/s-mal futtatta a Qwen3.6 35B-A3B-t \u2013 gyorsabban, mint az asztali DGX Spark. A b\u00f6kken\u0151 a Windows: a GPU a mem\u00f3ri\u00e1nak csak egy r\u00f6gz\u00edtett r\u00e9sz\u00e9t kapja \u2013 32&nbsp;GB-b\u00f3l legfeljebb 16&nbsp;GB-ot, 64&nbsp;GB-b\u00f3l 48&nbsp;GB-ot, 128&nbsp;GB-b\u00f3l pedig 96&nbsp;GB-ot.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Mem\u00f3ria:<\/strong> 32, 64 vagy 128&nbsp;GB \u00b7 <strong>s\u00e1vsz\u00e9less\u00e9g<\/strong> legfeljebb 300&nbsp;GB\/s<\/li>\n\n\n<li><strong>GPU:<\/strong> Blackwell, 6144 mag (teljes\u00edtm\u00e9nyben nagyj\u00e1b\u00f3l egy notebookos RTX 5070)<\/li>\n\n\n<li><strong>OS:<\/strong> Windows 11 on Arm, CUDA a 13.4-es verzi\u00f3t\u00f3l<\/li>\n\n\n<li><strong>Figyelem:<\/strong> az els\u0151 p\u00e9ld\u00e1nyok (Surface, HP) a tesztekben t\u00f6bb \u00f3r\u00e1nyi teljes terhel\u00e9s ut\u00e1n lefagytak; a Linux hivatalosan nem t\u00e1mogatott<\/li>\n\n\n<li><strong>Amerikai \u00e1r:<\/strong> 2599 doll\u00e1rt\u00f3l (Surface Laptop Ultra 24&nbsp;GB-tal) 5899,99 doll\u00e1rig (Surface 128&nbsp;GB-tal), 2026. okt\u00f3ber<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Kinek val\u00f3:<\/strong> nagy modellek Windows-notebookon, k\u00f3dol\u00f3 \u00fcgyn\u00f6k\u00f6k, kreat\u00edv munka. <strong>Kinek nem:<\/strong> a 32&nbsp;GB-os v\u00e1ltozat, hossz\u00fa tan\u00edt\u00e1sok, csak x86-ra el\u00e9rhet\u0151 szoftverek.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"ryzen-ai-max-395\">4. Ryzen AI Max+ 395 \u2013 a legolcs\u00f3bb 128&nbsp;GB<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A 16 magos Ryzennel \u00e9s Radeon 8060S grafik\u00e1val szerelt mini PC-kben 128&nbsp;GB k\u00f6z\u00f6s mem\u00f3ria van. A kimeneti sebess\u00e9g a DGX Spark\u00e9hoz hasonl\u00f3 (gpt-oss-120b 56,6 token\/s-mal) nagyj\u00e1b\u00f3l feleannyi p\u00e9nz\u00e9rt, \u00e9s Windowst, Linuxot \u2013 meg j\u00e1t\u00e9kokat \u2013 is futtatnak. A gyenge pont a promptbeolvas\u00e1s: k\u00e9tszer-h\u00e1romszor lassabb, mint a Sparkon, \u00e9s hossz\u00fa kontextusn\u00e1l tov\u00e1bb esik \u2013 a Vulkan-meghajt\u00f3val 146, ROCm-mel 371 token\/s-ra.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Mem\u00f3ria:<\/strong> 128&nbsp;GB LPDDR5x \u00b7 <strong>s\u00e1vsz\u00e9less\u00e9g<\/strong> 256&nbsp;GB\/s (m\u00e9rve ~215)<\/li>\n\n\n<li><strong>A GPU-nak:<\/strong> 96&nbsp;GB Windows alatt, ~124&nbsp;GB Linux alatt<\/li>\n\n\n<li><strong>Szoftver:<\/strong> Vulkan vagy ROCm (a gfx1151 2026 \u00f3ta hivatalosan t\u00e1mogatott)<\/li>\n\n\n<li><strong>A HP Z2 Mini G1a<\/strong> dr\u00e1g\u00e1bb, de ECC mem\u00f3ri\u00e1s, h\u00e1rom\u00e9ves garanci\u00e1s, min\u0151s\u00edtett munka\u00e1llom\u00e1s<\/li>\n\n\n<li><strong>Amerikai \u00e1rp\u00e9lda:<\/strong> a 128&nbsp;GB-os Framework Desktop 1999 doll\u00e1r\u00e9rt jelent meg, a 2026-os mem\u00f3ria\u00e1r-emel\u00e9s ut\u00e1n 2459 doll\u00e1rba ker\u00fcl<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Kinek val\u00f3:<\/strong> a legt\u00f6bb mem\u00f3ria a p\u00e9nz\u00e9rt, nagy MoE-modellek, cseveg\u00e9s \u00e9s \u00f6sszefoglal\u00f3k. <strong>Kinek nem:<\/strong> \u00fcgyn\u00f6k\u00f6k hatalmas kontextussal, CUDA, illetve a 32\u201364&nbsp;GB-os ROG Flow Z13 tablet nagy modellekhez.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"mac-studio-m5-ultra\">5. Mac Studio M5 Ultra \u2013 a mem\u00f3ria kir\u00e1lya<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Ak\u00e1r 512&nbsp;GB mem\u00f3ria \u00e9s 1,2&nbsp;TB\/s s\u00e1vsz\u00e9less\u00e9g. A MacStories tesztj\u00e9ben az M5 Ultra n\u00e9gyszer gyorsabban olvasta be a promptot, mint az el\u0151z\u0151 M3 Ultra, \u00e9s 50%-kal gyorsabban \u00edrt. A 256&nbsp;GB-os v\u00e1ltozatba m\u00e9g a Qwen3&nbsp;235B is belef\u00e9r. T\u00f6bb egys\u00e9g \u00f6sszekapcsolhat\u00f3 Thunderbolt 5-\u00f6n \u00e9s RDMA-n kereszt\u00fcl: az Apple szerint n\u00e9gy M5 Ultra 36,7 token\/s-mal futtat egy 816 milli\u00e1rd param\u00e9teres modellt.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Mem\u00f3ria:<\/strong> 96, 256 vagy 512&nbsp;GB (az 512&nbsp;GB-os okt\u00f3ber v\u00e9g\u00e9n \u00e9rkezik) \u00b7 <strong>s\u00e1vsz\u00e9less\u00e9g<\/strong> 1228&nbsp;GB\/s<\/li>\n\n\n<li><strong>Fogyaszt\u00e1s:<\/strong> 9&nbsp;W \u00fcresj\u00e1ratban, modellfuttat\u00e1s k\u00f6zben 300&nbsp;W k\u00f6r\u00fcl<\/li>\n\n\n<li><strong>Csatlakoz\u00f3k:<\/strong> 6\u00d7 Thunderbolt 5<\/li>\n\n\n<li><strong>Figyelem:<\/strong> a mem\u00f3ria ut\u00f3lag nem b\u0151v\u00edthet\u0151; szeptemberben a v\u00e1rakoz\u00e1si id\u0151 ak\u00e1r 18 h\u00e9t is volt<\/li>\n\n\n<li><strong>Amerikai lista\u00e1r:<\/strong> 5499 doll\u00e1rt\u00f3l (2026. augusztus); a 96 helyett 256&nbsp;GB 4000 doll\u00e1r fel\u00e1r<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Kinek val\u00f3:<\/strong> 100\u2013250B-s modellek, \u00fcgyn\u00f6k\u00f6k hossz\u00fa kontextussal, kis csapatok. <strong>Kinek nem:<\/strong> legfeljebb 30B-s modellek, CUDA, sz\u0171k k\u00f6lts\u00e9gvet\u00e9s.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"mac-mini-m5-pro\">6. Mac mini M5 Pro \u2013 az \u00e9sszer\u0171 bel\u00e9p\u0151<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A legolcs\u00f3bb Mac, amely val\u00f3ban hasznos modelleket futtat. 307&nbsp;GB\/s s\u00e1vsz\u00e9less\u00e9g, legfeljebb 64&nbsp;GB mem\u00f3ria \u00e9s Thunderbolt 5. A neur\u00e1lis gyors\u00edt\u00f3knak k\u00f6sz\u00f6nhet\u0151en nagyj\u00e1b\u00f3l h\u00e1romszor gyorsabban olvassa be a promptot, mint az el\u0151z\u0151 M4 Pro. A 64&nbsp;GB-os v\u00e1ltozat k\u00e9nyelmesen futtat 32 milli\u00e1rd param\u00e9terig terjed\u0151 modelleket; a s\u0171r\u0171 Llama 70B belef\u00e9r, de csak k\u00f6r\u00fclbel\u00fcl 7 token\/s-ra k\u00e9pes, a gpt-oss-120b pedig nem f\u00e9r bele.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Mem\u00f3ria:<\/strong> 24, 48 vagy 64&nbsp;GB \u00b7 <strong>s\u00e1vsz\u00e9less\u00e9g<\/strong> 307&nbsp;GB\/s<\/li>\n\n\n<li><strong>Fogyaszt\u00e1s:<\/strong> 6&nbsp;W \u00fcresj\u00e1ratban, max. 145&nbsp;W \u00b7 \u00fcresj\u00e1ratban gyakorlatilag hangtalan<\/li>\n\n\n<li><strong>Amerikai lista\u00e1r:<\/strong> 1699 doll\u00e1rt\u00f3l (alapkonfigur\u00e1ci\u00f3, 2026. augusztus)<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Kinek val\u00f3:<\/strong> 20\u201335B-s modellek, csendes otthoni AI-szerver. <strong>Kinek nem:<\/strong> gpt-oss-120b, s\u0171r\u0171 70B, a 24&nbsp;GB-os v\u00e1ltozat.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"mac-mini-m6\">7. Mac mini M6 \u2013 olcs\u00f3, de nem 16&nbsp;GB-tal<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A legolcs\u00f3bb Mac neur\u00e1lis gyors\u00edt\u00f3kkal. 32&nbsp;GB-n\u00e1l \u00e9s 170&nbsp;GB\/s-n\u00e1l meg\u00e1ll \u2013 a 16&nbsp;GB-os v\u00e1ltozat s\u00e1vsz\u00e9less\u00e9ge csak 153&nbsp;GB\/s, \u00e9s a modellnek k\u00f6r\u00fclbel\u00fcl 12&nbsp;GB-ot hagy. Az el\u0151z\u0151, M4-es Mac minihez k\u00e9pest 4,5-szer gyorsabban olvassa be a promptot \u00e9s 50%-kal gyorsabban \u00edr, gyakorlatilag ugyanazon az \u00e1ron, ez\u00e9rt a r\u00e9gi modellt akci\u00f3san se vedd meg.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Mem\u00f3ria:<\/strong> 16, 24 vagy 32&nbsp;GB (maximum) \u00b7 <strong>s\u00e1vsz\u00e9less\u00e9g<\/strong> 153 \/ 170&nbsp;GB\/s<\/li>\n\n\n<li><strong>Fogyaszt\u00e1s:<\/strong> 4&nbsp;W \u00fcresj\u00e1ratban, max. 70&nbsp;W<\/li>\n\n\n<li><strong>Figyelem:<\/strong> csak Thunderbolt 4, \u00edgy t\u00f6bb egys\u00e9g nem kapcsolhat\u00f3 \u00f6ssze<\/li>\n\n\n<li><strong>Amerikai lista\u00e1r:<\/strong> 899 doll\u00e1rt\u00f3l (alapkonfigur\u00e1ci\u00f3, 2026. augusztus)<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Kinek val\u00f3:<\/strong> 7\u201314B-s modellek, gpt-oss-20b, Qwen3 30B-A3B. <strong>Kinek nem:<\/strong> a 16&nbsp;GB-os v\u00e1ltozat, 70B-s \u00e9s nagyobb modellek.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"macbook-pro-m5-max\">8. MacBook Pro M5 Max \u2013 AI \u00fatk\u00f6zben<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Ugyanaz a chip, mint a Mac Studi\u00f3ban, \u00edgy ugyanaz a sebess\u00e9g is: ez a leggyorsabb notebook helyi modellekhez. A k\u00f6zepes m\u00e9ret\u0171 Qwen3.6 35B-A3B-vel 90 token\/s felett gener\u00e1l. A 14 h\u00fcvelykes v\u00e1ltozat azonban tart\u00f3s terhel\u00e9s alatt visszafogja mag\u00e1t (96-r\u00f3l 42&nbsp;W-ra), a t\u00f6lt\u00e9s pedig 97&nbsp;W-ban maxim\u00e1lt. Egy azonos chipes, 128&nbsp;GB-os Mac Studio j\u00f3val olcs\u00f3bb.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Mem\u00f3ria:<\/strong> 36, 48, 64 vagy 128&nbsp;GB \u00b7 <strong>s\u00e1vsz\u00e9less\u00e9g<\/strong> 460 \/ 614&nbsp;GB\/s<\/li>\n\n\n<li><strong>Akkumul\u00e1tor:<\/strong> 72,4&nbsp;Wh (14&#8243;) \/ 100&nbsp;Wh (16&#8243;)<\/li>\n\n\n<li><strong>Amerikai lista\u00e1r:<\/strong> 3599 doll\u00e1rt\u00f3l (14&#8243;) \u00e9s 3899 doll\u00e1rt\u00f3l (16&#8243;), 2026. m\u00e1rcius; a 64 helyett 128&nbsp;GB 1000 doll\u00e1r fel\u00e1r<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Kinek val\u00f3:<\/strong> akinek \u00fatk\u00f6zben is nagy modellek kellenek, ide\u00e1lisan a 16 h\u00fcvelykes, 64\u2013128&nbsp;GB-os v\u00e1ltozat. <strong>Kinek nem:<\/strong> a 36&nbsp;GB-os v\u00e1ltozat, \u00e9jjel-nappali terhel\u00e9s, aki asztaln\u00e1l dolgozik.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"rtx-5090\">9. RTX 5090 \u2013 a leggyorsabb, de csak 32&nbsp;GB<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">1792&nbsp;GB\/s s\u00e1vsz\u00e9less\u00e9g: a gpt-oss-20b k\u00f6zel 300 token\/s-mal \u00edr, a k\u00e1rtya pedig nagyj\u00e1b\u00f3l k\u00e9tszer olyan gyorsan olvassa be a hossz\u00fa promptot, mint egy M5 Ultr\u00e1s Mac Studio. A 32&nbsp;GB 4 biten nagyj\u00e1b\u00f3l 32 milli\u00e1rd param\u00e9teres modelleket fogad be, m\u00e9g hossz\u00fa kontextussal is; egy 70 milli\u00e1rd param\u00e9teres modell nem f\u00e9r bele. A nagy MoE-modellek r\u00e9szben kiszervezhet\u0151k a rendszermem\u00f3ri\u00e1ba \u2013 a gpt-oss-120b \u00edgy k\u00f6r\u00fclbel\u00fcl 30 token\/s-mal fut.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Mem\u00f3ria:<\/strong> 32&nbsp;GB GDDR7 \u00b7 <strong>s\u00e1vsz\u00e9less\u00e9g<\/strong> 1792&nbsp;GB\/s<\/li>\n\n\n<li><strong>Fogyaszt\u00e1s:<\/strong> 575&nbsp;W, 1000&nbsp;W-os t\u00e1p aj\u00e1nlott<\/li>\n\n\n<li><strong>Amerikai lista\u00e1r:<\/strong> 1999 doll\u00e1r a megjelen\u00e9skor (2025. janu\u00e1r); a bolti \u00e1rak 2026-ban a mem\u00f3riahi\u00e1ny miatt j\u00f3val magasabbak<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Kinek val\u00f3:<\/strong> k\u00f3dol\u00f3 \u00fcgyn\u00f6k\u00f6k \u00e9s RAG legfeljebb 32B-s modellekkel, CUDA-fejleszt\u00e9s, j\u00e1t\u00e9k. <strong>Kinek nem:<\/strong> 70B-s \u00e9s nagyobb modellek, csendes, alacsony fogyaszt\u00e1s\u00fa szerver.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"rtx-5060-ti-16gb\">10. RTX 5060 Ti 16&nbsp;GB \u2013 a legolcs\u00f3bb \u00e9rtelmes bel\u00e9p\u0151<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A legolcs\u00f3bb \u00faj, 16&nbsp;GB-os videok\u00e1rtya. A gpt-oss-20b-t k\u00f6r\u00fclbel\u00fcl 92, a Qwen3 8B-t 69, a Qwen3 14B-t 41 token\/s-mal futtatja \u2013 gyorsabban, mint ahogy olvasni tudsz. A mem\u00f3ria-s\u00e1vsz\u00e9less\u00e9ge viszont fele az RTX 5070 Ti-\u00e9nak, \u00edgy nagyj\u00e1b\u00f3l feleakkora temp\u00f3ban \u00edr, \u00e9s a 30 milli\u00e1rd param\u00e9ter k\u00f6r\u00fcli modellek nem f\u00e9rnek bele a 16&nbsp;GB-ba.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Mem\u00f3ria:<\/strong> 16&nbsp;GB GDDR7 \u00b7 <strong>s\u00e1vsz\u00e9less\u00e9g<\/strong> 448&nbsp;GB\/s<\/li>\n\n\n<li><strong>Fogyaszt\u00e1s:<\/strong> 180&nbsp;W, egy 600&nbsp;W-os t\u00e1p el\u00e9g<\/li>\n\n\n<li><strong>Amerikai lista\u00e1r:<\/strong> 429 doll\u00e1r a 16&nbsp;GB-os v\u00e1ltozat\u00e9rt a megjelen\u00e9skor (2025. \u00e1prilis)<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Kinek val\u00f3:<\/strong> els\u0151 komoly k\u00eds\u00e9rletek, 8\u201320B-s modellek, k\u00f6lts\u00e9gk\u00edm\u00e9l\u0151 AI PC. <strong>Kinek nem:<\/strong> 27B-s \u00e9s nagyobb modellek.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"rtx-5070-ti-vs-5080\">11. RTX 5070 Ti vs. RTX 5080 \u2013 AI-ra szinte egyform\u00e1k<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Mindk\u00e9t k\u00e1rty\u00e1n 16&nbsp;GB mem\u00f3ria \u00e9s szinte ugyanakkora s\u00e1vsz\u00e9less\u00e9g (896 vs. 960&nbsp;GB\/s) van, \u00edgy k\u00f6zel azonos sebess\u00e9ggel \u00edrnak: a szabv\u00e1nyos llama.cpp-tesztben 182,4 vs. 184,7 token\/s-mal, a Qwen3 14B-n 74 vs. 81 token\/s-mal. Pontosan ugyanazok a modellek f\u00e9rnek bel\u00e9j\u00fck. Az 5080 f\u0151leg j\u00e1t\u00e9kokban gyorsabb \u2013 \u00e9s a megjelen\u00e9si \u00e1ra 250 doll\u00e1rral magasabb volt (999 vs. 749 doll\u00e1r).<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Tan\u00e1csunk:<\/strong> AI-ra vedd az RTX 5070 Ti-t, a megsp\u00f3rolt p\u00e9nzt pedig ford\u00edtsd nagyobb SSD-re vagy t\u00f6bb RAM-ra. RTX 5080-at csak akkor vegy\u00e9l, ha f\u0151leg j\u00e1tszani fogsz rajta.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"rtx-pro-6000\">12. RTX PRO 6000 \u2013 96&nbsp;GB egyetlen k\u00e1rty\u00e1n<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Az egyetlen k\u00e1rtya, amelybe teljes eg\u00e9sz\u00e9ben belef\u00e9r a Llama 3.3 70B (32 token\/s), a gpt-oss-120b (170,6 token\/s a Max-Q v\u00e1ltozatban) \u00e9s a Qwen3.5&nbsp;122B (k\u00f6r\u00fclbel\u00fcl 98 token\/s). A CUDA-nak \u00e9s a vLLM-nek k\u00f6sz\u00f6nhet\u0151en egyszerre t\u00f6bb felhaszn\u00e1l\u00f3t is kiszolg\u00e1lhat. \u00d6n\u00e1ll\u00f3 k\u00e1rtyak\u00e9nt \u00e9s el\u0151re \u00f6sszeszerelt munka\u00e1llom\u00e1sban is megv\u00e1s\u00e1rolhat\u00f3.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Mem\u00f3ria:<\/strong> 96&nbsp;GB GDDR7, ECC-vel \u00b7 <strong>s\u00e1vsz\u00e9less\u00e9g<\/strong> 1792&nbsp;GB\/s<\/li>\n\n\n<li><strong>Fogyaszt\u00e1s:<\/strong> 300&nbsp;W (Max-Q); a teljes, 600&nbsp;W-os v\u00e1ltozat k\u00f6r\u00fclbel\u00fcl 13%-kal gyorsabban \u00edr<\/li>\n\n\n<li><strong>Amerikai lista\u00e1r:<\/strong> 8565 doll\u00e1r a megjelen\u00e9skor; az NVIDIA k\u00e9s\u0151bb 13&nbsp;250 doll\u00e1rra emelte<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Kinek val\u00f3:<\/strong> c\u00e9gek \u00e9s csapatok, amelyek h\u00e1zon bel\u00fcl akarnak nagy modelleket futtatni vagy tan\u00edtani. <strong>Kinek nem:<\/strong> mag\u00e1nszem\u00e9lyek \u00e9s hobbist\u00e1k.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"gb10-klony\">13. DGX Spark az ASUS-t\u00f3l, MSI-t\u00f3l, Lenov\u00f3t\u00f3l, Dellt\u0151l \u00e9s Gigabyte-t\u00f3l \u2013 melyiket v\u00e1laszd<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u00d6t m\u00e1sik gy\u00e1rt\u00f3 is ugyanazzal a GB10 chippel \u00e1rul g\u00e9pet, mint amilyen a DGX Sparkban van. Az AI-teljes\u00edtm\u00e9ny csak n\u00e9h\u00e1ny sz\u00e1zal\u00e9kban t\u00e9r el; a k\u00fcl\u00f6nbs\u00e9gek a t\u00e1rhelyben, a t\u00e1pegys\u00e9gben, a h\u0171t\u00e9sben \u00e9s a t\u00e1mogat\u00e1sban vannak \u2013 no meg az \u00e1rban.<\/p>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table class=\"has-fixed-layout\"><thead><tr><th>Modell<\/th><th>T\u00e1rhely<\/th><th>T\u00e1pegys\u00e9g<\/th><th>Kiemelked\u0151 jellemz\u0151<\/th><\/tr><\/thead><tbody><tr><td>NVIDIA \/ PNY DGX Spark<\/td><td>4&nbsp;TB, \u00f6ntitkos\u00edt\u00f3<\/td><td>240&nbsp;W<\/td><td>az eredeti<\/td><\/tr><tr><td>GIGABYTE AI TOP ATOM<\/td><td>4&nbsp;TB PCIe 5.0<\/td><td>240&nbsp;W<\/td><td>leggyorsabb SSD<\/td><\/tr><tr><td>ASUS Ascent GX10<\/td><td>1&nbsp;TB PCIe 4.0 \/ 4&nbsp;TB PCIe 5.0<\/td><td>240&nbsp;W<\/td><td>el\u00f6l elhelyezett bekapcsol\u00f3gomb<\/td><\/tr><tr><td>Lenovo ThinkStation PGX<\/td><td>1 \/ 2 \/ 4&nbsp;TB, \u00f6ntitkos\u00edt\u00f3<\/td><td>240&nbsp;W<\/td><td>egy\u00e9ves helysz\u00edni szerviz<\/td><\/tr><tr><td>Dell Pro Max with GB10<\/td><td>PCIe 4.0 (a 2&nbsp;TB-os QLC)<\/td><td>280&nbsp;W<\/td><td>er\u0151sebb t\u00e1p, legforr\u00f3bb SSD<\/td><\/tr><tr><td>MSI EdgeXpert<\/td><td>4&nbsp;TB<\/td><td>\u2013<\/td><td>\u00faj, m\u00e9g nincs f\u00fcggetlen teszt<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Tan\u00e1csunk:<\/strong> \u00e9sszer\u0171 p\u00e9nz\u00e9rt a legjobb felszerelts\u00e9get a Gigabyte AI TOP ATOM adja 4&nbsp;TB gyors t\u00e1rhellyel; a legolcs\u00f3bb bev\u00e1lt kiindul\u00f3pont az 1&nbsp;TB-os ASUS Ascent GX10, amely az USA-ban nagyj\u00e1b\u00f3l 3000 doll\u00e1r\u00e9rt jelent meg. Ha k\u00e9t egys\u00e9get akarsz \u00f6sszekapcsolni, QSFP-k\u00e1belre lesz sz\u00fcks\u00e9ged. 2026. okt\u00f3ber 23. \u00f3ta \u00fajdons\u00e1g: a GB10 64&nbsp;GB-os v\u00e1ltozata, amelyet csak a partnerm\u00e1rk\u00e1k \u00e1rulnak, 4999 doll\u00e1rt\u00f3l.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"radeon-ai-pro-r9700\">14. Radeon AI PRO R9700 \u2013 32&nbsp;GB az RTX 5090 \u00e1r\u00e1nak t\u00f6red\u00e9k\u00e9\u00e9rt<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A legolcs\u00f3bb \u00faj k\u00e1rtya 32&nbsp;GB-tal. Ugyanabban a Vulkan-tesztben k\u00f6r\u00fclbel\u00fcl 40%-kal lassabban \u00edr, mint az RTX 5090, de ugyanazok a modellek f\u00e9rnek bele \u2013 nagyj\u00e1b\u00f3l 30 milli\u00e1rd param\u00e9terig, ak\u00e1r hosszabb kontextussal is. A gpt-oss-20b \u00e9s a Qwen3 30B-A3B 150 token\/s felett fut rajta. \u00d3vatosan a marketinggel: az \u201eak\u00e1r \u00f6tsz\u00f6r gyorsabb, mint az RTX 5080\u201d \u00e1ll\u00edt\u00e1s csak olyan modellre igaz, amely nem f\u00e9r bele 16&nbsp;GB-ba.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Mem\u00f3ria:<\/strong> 32&nbsp;GB GDDR6 \u00b7 <strong>s\u00e1vsz\u00e9less\u00e9g<\/strong> ~640&nbsp;GB\/s \u00b7 <strong>fogyaszt\u00e1s<\/strong> 300&nbsp;W<\/li>\n\n\n<li><strong>Szoftver:<\/strong> Vulkan (\u00e1ltal\u00e1ban gyorsabb kimenet) vagy ROCm (gyorsabb promptbeolvas\u00e1s); CUDA nincs<\/li>\n\n\n<li><strong>Amerikai lista\u00e1r:<\/strong> 1299 doll\u00e1r (2025. okt\u00f3ber)<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Kinek val\u00f3:<\/strong> 27\u201332B-s modellek hossz\u00fa kontextussal, t\u00f6bb k\u00e1rtya egy PC-ben. <strong>Kinek nem:<\/strong> csak CUDA-val m\u0171k\u00f6d\u0151 eszk\u00f6z\u00f6k, maxim\u00e1lis sebess\u00e9g.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"ai-pc-copilot\">15. \u201eAI PC-k\u201d \u00e9s Copilot+ notebookok \u2013 egy m\u00edtosz<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A Copilot+ a Microsoft tan\u00fas\u00edtv\u00e1nya a Windows-funkci\u00f3khoz \u2013 kameraeffektusok, \u00e9l\u0151 feliratoz\u00e1s, Recall vagy a be\u00e9p\u00edtett kis Phi Silica modell. A nagy nyelvi modellek futtat\u00e1s\u00e1nak k\u00e9pess\u00e9g\u00e9r\u0151l semmit nem \u00e1rul el. Sem az Ollama, sem az LM Studio nem haszn\u00e1lja a neur\u00e1lis feldolgoz\u00f3egys\u00e9get (NPU-t), a llama.cpp pedig csak k\u00eds\u00e9rleti jelleggel; az NPU-t csak gy\u00e1rt\u00f3i eszk\u00f6z\u00f6k haszn\u00e1lj\u00e1k, p\u00e9ld\u00e1ul a Microsoft Foundry Local vagy az AMD Lemonade, kiv\u00e1lasztott modellekkel.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><a href=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.hu_-scaled.webp\"><img loading=\"lazy\" decoding=\"async\" width=\"2560\" height=\"1259\" src=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.hu_-scaled.webp\" alt=\"Egy 8 milli\u00e1rd param\u00e9teres modell kimeneti sebess\u00e9ge: Copilot+ notebook \u00e9s videok\u00e1rtya\" class=\"wp-image-29370\" srcset=\"https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.hu_-scaled.webp 2560w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.hu_-300x148.webp 300w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.hu_-1024x504.webp 1024w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.hu_-768x378.webp 768w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.hu_-1536x756.webp 1536w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.hu_-2048x1007.webp 2048w, https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/ai-pc-npu.hu_-610x300.webp 610w\" sizes=\"auto, (max-width: 2560px) 100vw, 2560px\" \/><\/a><figcaption class=\"wp-element-caption\">Nyelvi modellekn\u00e9l m\u00e9g egy olcs\u00f3 videok\u00e1rtya is t\u00f6bbsz\u00f6r\u00f6sen gyorsabb az NPU-n\u00e1l.<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">R\u00e1ad\u00e1sul ezekben a notebookokban k\u00f6r\u00fclbel\u00fcl 135&nbsp;GB\/s s\u00e1vsz\u00e9less\u00e9g\u0171 mem\u00f3ria \u00e9s 16&nbsp;GB RAM van, ami nagyj\u00e1b\u00f3l 8&nbsp;GB-ot hagy a modellnek. 4 milli\u00e1rd param\u00e9terig terjed\u0151 modellekhez \u00e9s a Windows-funkci\u00f3khoz megfelelnek \u2013 a \u201esaj\u00e1t ChatGPT otthon\u201d \u00e9lm\u00e9nyhez nem.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Kinek val\u00f3:<\/strong> irodai munka, hossz\u00fa akkumul\u00e1tor-\u00fczemid\u0151, kis 1\u20134B-s modellek. <strong>Kinek nem:<\/strong> aki 14B-s vagy nagyobb modelleket akar futtatni.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"spotreba\">Fogyaszt\u00e1s \u00e9s \u00fczemeltet\u00e9si k\u00f6lts\u00e9gek<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Ha a g\u00e9p \u00e9jjel-nappal otthoni AI-szerverk\u00e9nt fog futni, az \u00e1ramfogyaszt\u00e1s \u00e9s a zaj is sz\u00e1m\u00edt.<\/p>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table class=\"has-fixed-layout\"><thead><tr><th>Eszk\u00f6z<\/th><th>\u00dcresj\u00e1rat<\/th><th>Modellfuttat\u00e1s \/ maximum<\/th><th>Megjegyz\u00e9s<\/th><\/tr><\/thead><tbody><tr><td>Mac mini M6<\/td><td>4&nbsp;W<\/td><td>max. 70&nbsp;W<\/td><td>gyakorlatilag hangtalan<\/td><\/tr><tr><td>Mac mini M5 Pro<\/td><td>6&nbsp;W<\/td><td>max. 145&nbsp;W<\/td><td>5&nbsp;dBA \u00fcresj\u00e1ratban (Apple)<\/td><\/tr><tr><td>Mac Studio M5 Max<\/td><td>7&nbsp;W<\/td><td>max. 200&nbsp;W<\/td><td>csendes<\/td><\/tr><tr><td>Mac Studio M5 Ultra<\/td><td>9&nbsp;W<\/td><td>~300&nbsp;W AI futtat\u00e1sakor<\/td><td>a tesztek szerint csendes<\/td><\/tr><tr><td>DGX Spark<\/td><td>25\u201345&nbsp;W<\/td><td>60\u201390&nbsp;W AI futtat\u00e1sakor, cs\u00facsban ak\u00e1r 200&nbsp;W<\/td><td>a firmware verzi\u00f3j\u00e1t\u00f3l f\u00fcgg<\/td><\/tr><tr><td>Ryzen AI Max+ 395 mini PC<\/td><td>~13&nbsp;W<\/td><td>120\u2013160&nbsp;W<\/td><td>a HP Z2 Mini terhel\u00e9s alatt hangos<\/td><\/tr><tr><td>RTX 5090 (csak a k\u00e1rtya)<\/td><td>\u2013<\/td><td>legfeljebb 575&nbsp;W<\/td><td>plusz a PC t\u00f6bbi r\u00e9sze<\/td><\/tr><tr><td>RTX 5060 Ti (csak a k\u00e1rtya)<\/td><td>\u2013<\/td><td>legfeljebb 180&nbsp;W<\/td><td><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Az adatok gy\u00e1rt\u00f3kt\u00f3l \u00e9s tesztekb\u0151l (ServeTheHome, Tom&#8217;s Hardware, Notebookcheck) sz\u00e1rmaznak, \u00e9s nem azonos m\u00f3dszerrel m\u00e9rt\u00e9k \u0151ket \u2013 tekintsd \u0151ket t\u00e1j\u00e9koztat\u00f3 jelleg\u0171nek.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"software\">Szoftver: mivel futtasd a modellt<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Ollama<\/strong> \u2013 a legk\u00f6nnyebb kezd\u00e9s minden platformon; egyetlen paranccsal let\u00f6ltheted a modellt. Macen a 0.19-es verzi\u00f3t\u00f3l (egyel\u0151re el\u0151zetes v\u00e1ltozatban, kiv\u00e1lasztott modellekn\u00e9l) az MLX-et \u00e9s az M5 neur\u00e1lis gyors\u00edt\u00f3it haszn\u00e1lja.<\/li>\n\n\n<li><strong>LM Studio<\/strong> \u2013 \u00e1ttekinthet\u0151 alkalmaz\u00e1s grafikus fel\u00fclettel. <strong>Tipp M5-\u00f6s Macekhez:<\/strong> GGUF-modellekn\u00e9l m\u00e9g nem haszn\u00e1lja a neur\u00e1lis gyors\u00edt\u00f3kat, \u00e9s nagyj\u00e1b\u00f3l feleakkora sebess\u00e9ggel olvassa be a promptot; haszn\u00e1lj helyett\u00fck MLX form\u00e1tum\u00fa modelleket.<\/li>\n\n\n<li><strong>llama.cpp<\/strong> \u2013 a legt\u00f6bb m\u00e1s eszk\u00f6z alapja, a legt\u00f6bb be\u00e1ll\u00edt\u00e1si lehet\u0151s\u00e9ggel \u00e9s az \u00faj hardverek leggyorsabb t\u00e1mogat\u00e1s\u00e1val. Ryzen AI Maxon pr\u00f3b\u00e1ld ki mindk\u00e9t backendet: a Vulkan \u00e1ltal\u00e1ban gyorsabban \u00edr, a ROCm gyorsabban olvassa be a hossz\u00fa promptot.<\/li>\n\n\n<li><strong>MLX<\/strong> \u2013 az Apple k\u00f6nyvt\u00e1ra, a legjobb teljes\u00edtm\u00e9nnyel Macen, bele\u00e9rtve t\u00f6bb Mac \u00f6sszekapcsol\u00e1s\u00e1t Thunderbolt 5-\u00f6n kereszt\u00fcl.<\/li>\n\n\n<li><strong>vLLM, SGLang, TensorRT-LLM<\/strong> \u2013 szervereszk\u00f6z\u00f6k NVIDIA-hoz, amikor egy modell egyszerre t\u00f6bb felhaszn\u00e1l\u00f3t szolg\u00e1l ki.<\/li>\n\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"chyby\">A leggyakoribb v\u00e1s\u00e1rl\u00e1si hib\u00e1k<\/h2>\n\n\n\n<ol class=\"wp-block-list\">\n\n<li><strong>36&nbsp;GB-os Mac \u201eAI-ra\u201d.<\/strong> Gyeng\u00e9bb a chipje, \u00e9s a modellnek csak k\u00f6r\u00fclbel\u00fcl 27&nbsp;GB marad.<\/li>\n\n\n<li><strong>RTX Spark notebook 32&nbsp;GB-tal.<\/strong> A Windows legfeljebb 16&nbsp;GB-ot ad a GPU-nak.<\/li>\n\n\n<li><strong>Mac mini M6 16&nbsp;GB-tal.<\/strong> Lassabb mem\u00f3ria, \u00e9s a modellnek csak k\u00f6r\u00fclbel\u00fcl 12&nbsp;GB jut.<\/li>\n\n\n<li><strong>Copilot+ notebook helyi AI-ra.<\/strong> A h\u00e9tk\u00f6znapi eszk\u00f6z\u00f6k nem haszn\u00e1lj\u00e1k az NPU-t.<\/li>\n\n\n<li><strong>RTX 5080 az 5070 Ti helyett, csak AI-ra.<\/strong> Ugyanaz a 16&nbsp;GB, szinte ugyanaz a sebess\u00e9g.<\/li>\n\n\n<li><strong>A mem\u00f3ria-s\u00e1vsz\u00e9less\u00e9g figyelmen k\u00edv\u00fcl hagy\u00e1sa.<\/strong> A nagy, de lass\u00fa mem\u00f3ria s\u0171r\u0171 modellekn\u00e9l lass\u00fa kimenetet jelent.<\/li>\n\n\n<li><strong>Sp\u00f3rol\u00e1s a mem\u00f3ri\u00e1n.<\/strong> A Maceken, a Ryzen AI Maxon \u00e9s az RTX Sparkon a mem\u00f3ria be van forrasztva, ut\u00f3lag nem b\u0151v\u00edthet\u0151.<\/li>\n\n<\/ol>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"faq\">Gyakori k\u00e9rd\u00e9sek<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>El\u00e9g a 16&nbsp;GB a helyi AI-hoz?<\/strong> 14\u201320 milli\u00e1rd param\u00e9teres modellekig igen \u2013 p\u00e9ld\u00e1ul a gpt-oss-20b-hez vagy a Qwen3 14B-hez. A 30 milli\u00e1rd k\u00f6r\u00fcli modellekhez 24\u201332&nbsp;GB, a 70 milli\u00e1rdosokhoz legal\u00e1bb 48&nbsp;GB kell.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Mac vagy NVIDIA?<\/strong> A Mac gyorsabban \u00edr, \u00e9s t\u00f6bb mem\u00f3ri\u00e1t ad a p\u00e9nz\u00e9rt; az NVIDIA gyorsabban olvassa be a promptot, \u00e9s ott a CUDA. Cseveg\u00e9shez \u00e9s MoE-modellekkel dolgoz\u00f3 \u00fcgyn\u00f6k\u00f6kh\u00f6z a Mac j\u00f6n ki jobban, fejleszt\u00e9shez, tan\u00edt\u00e1shoz \u00e9s t\u00f6bb felhaszn\u00e1l\u00f3 kiszolg\u00e1l\u00e1s\u00e1hoz az NVIDIA.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Mit jelent a \u201eMoE\u201d?<\/strong> Mixture of experts, azaz szak\u00e9rt\u0151kever\u00e9k: a modellnek sok param\u00e9tere van, de minden sz\u00f3hoz csak egy kis r\u00e9sz\u00fcket haszn\u00e1lja. Sok mem\u00f3ria kell neki, de egy kisebb modell sebess\u00e9g\u00e9vel \u00edr.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Milyen gyorsan kell \u00edrnia a modellnek?<\/strong> K\u00f6r\u00fclbel\u00fcl 15 token\/s felel meg az olvas\u00e1si sebess\u00e9gnek. G\u00f6rd\u00fcl\u00e9keny cseveg\u00e9shez 30 vagy t\u00f6bb a k\u00e9nyelmes; sok sz\u00f6veget gener\u00e1l\u00f3 k\u00f3dol\u00f3 \u00fcgyn\u00f6k\u00f6kn\u00e9l 50 vagy t\u00f6bb a jobb.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Olcs\u00f3bb a helyi AI, mint az el\u0151fizet\u00e9s?<\/strong> \u00c1ltal\u00e1ban nem \u2013 a legjobb felh\u0151szolg\u00e1ltat\u00e1sokhoz m\u00e9rhet\u0151 modellekhez sok ezer doll\u00e1ros g\u00e9pek kellenek. A f\u0151 \u00e9rvek az adatv\u00e9delem, az adatok feletti kontroll \u00e9s a f\u00fcggetlens\u00e9g.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u00d6sszekapcsolhatok t\u00f6bb sz\u00e1m\u00edt\u00f3g\u00e9pet?<\/strong> Igen. K\u00e9t DGX Spark QSFP-k\u00e1bellel k\u00f6thet\u0151 \u00f6ssze, a Macek Thunderbolt 5-\u00f6n, RDMA-val (MLX vagy exo seg\u00edts\u00e9g\u00e9vel). A Mac mini M6 erre nem k\u00e9pes, mert csak Thunderbolt 4-e van.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"zdroje\">M\u00f3dszertan \u00e9s forr\u00e1sok<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Az eszk\u00f6z\u00f6ket nem tesztelt\u00fck saj\u00e1t magunk. A sz\u00e1mok ellen\u0151rz\u00f6tt forr\u00e1sokb\u00f3l sz\u00e1rmaznak, \u00e9s a grafikonokon mindig megadjuk \u0151ket. A sebess\u00e9get egyetlen programban (llama.cpp) hasonl\u00edtjuk \u00f6ssze, mert egy m\u00e1sik program vagy kvant\u00e1l\u00e1s k\u00f6nnyen 20\u201350%-kal eltolhatja az eredm\u00e9nyeket. A k\u00f6z\u00f6ss\u00e9gi m\u00e9r\u00e9seket (GitHub, f\u00f3rumok) k\u00fcl\u00f6n jel\u00f6lj\u00fck. Az \u00e1rak a gy\u00e1rt\u00f3i bejelent\u00e9sekb\u0151l \u00e9s tesztekb\u0151l vett, d\u00e1tummal ell\u00e1tott amerikai lista\u00e1rak.<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n\n<li>A modell kiszorul\u00e1sa a RAM-ba, RTX 5060 Ti: <a href=\"https:\/\/inventivehq.com\/blog\/vram-offload-cliff-gpu-layers-benchmark\" rel=\"noopener\" target=\"_blank\">inventivehq.com\/blog\/vram-offload-cliff-gpu-layers-benchmark<\/a><\/li>\n\n\n<li>DGX Spark benchmark: <a href=\"https:\/\/github.com\/ggml-org\/llama.cpp\/blob\/master\/benches\/dgx-spark\/dgx-spark.md\" rel=\"noopener\" target=\"_blank\">github.com\/ggml-org\/llama.cpp\/blob\/master\/benches\/dgx-spark\/dgx-spark.md<\/a><\/li>\n\n\n<li>gpt-oss \u00fatmutat\u00f3 \u00e9s GPU-m\u00e9r\u00e9sek: <a href=\"https:\/\/github.com\/ggml-org\/llama.cpp\/discussions\/15396\" rel=\"noopener\" target=\"_blank\">github.com\/ggml-org\/llama.cpp\/discussions\/15396<\/a><\/li>\n\n\n<li>M5 Max \u00e9s LM Studio: <a href=\"https:\/\/github.com\/lmstudio-ai\/lmstudio-bug-tracker\/issues\/2040\" rel=\"noopener\" target=\"_blank\">github.com\/lmstudio-ai\/lmstudio-bug-tracker\/issues\/2040<\/a><\/li>\n\n\n<li>Ryzen AI Max (Strix Halo): <a href=\"https:\/\/strix-halo-toolboxes.com\" rel=\"noopener\" target=\"_blank\">strix-halo-toolboxes.com<\/a>, <a href=\"https:\/\/llm-tracker.info\/_TOORG\/Strix-Halo\" rel=\"noopener\" target=\"_blank\">llm-tracker.info\/_TOORG\/Strix-Halo<\/a><\/li>\n\n\n<li>Apple: <a href=\"https:\/\/apple.com\/mac-studio\/specs\" rel=\"noopener\" target=\"_blank\">apple.com\/mac-studio\/specs<\/a>, <a href=\"https:\/\/apple.com\/mac-mini\/specs\" rel=\"noopener\" target=\"_blank\">apple.com\/mac-mini\/specs<\/a>, <a href=\"https:\/\/machinelearning.apple.com\/research\/mlx-distributed-inference\" rel=\"noopener\" target=\"_blank\">machinelearning.apple.com\/research\/mlx-distributed-inference<\/a><\/li>\n\n\n<li>M5 Ultra: <a href=\"https:\/\/macstories.net\/stories\/m5-ultra-mac-studio-review-the-dream-mac-for-local-ai-agents\" rel=\"noopener\" target=\"_blank\">macstories.net\/stories\/m5-ultra-mac-studio-review-the-dream-mac-for-local-ai-agents<\/a><\/li>\n\n\n<li>RTX Spark: <a href=\"https:\/\/theregister.com\" rel=\"noopener\" target=\"_blank\">theregister.com<\/a> (2026. okt\u00f3ber 16.), <a href=\"https:\/\/servethehome.com\" rel=\"noopener\" target=\"_blank\">servethehome.com<\/a>, <a href=\"https:\/\/storagereview.com\" rel=\"noopener\" target=\"_blank\">storagereview.com<\/a><\/li>\n\n\n<li>GeForce videok\u00e1rty\u00e1k: <a href=\"https:\/\/nvidia.com\/en-us\/geforce\/graphics-cards\/compare\" rel=\"noopener\" target=\"_blank\">nvidia.com\/en-us\/geforce\/graphics-cards\/compare<\/a>, <a href=\"https:\/\/hardware-corner.net\/gpu-llm-benchmarks\" rel=\"noopener\" target=\"_blank\">hardware-corner.net\/gpu-llm-benchmarks<\/a><\/li>\n\n\n<li>llama.cpp ranglist\u00e1k: <a href=\"https:\/\/github.com\/ggml-org\/llama.cpp\/discussions\/15013\" rel=\"noopener\" target=\"_blank\">github.com\/ggml-org\/llama.cpp\/discussions\/15013<\/a> \u00e9s \/10879<\/li>\n\n\n<li>Copilot+ PC: <a href=\"https:\/\/microsoft.com\/windows\/windows-11-specifications\" rel=\"noopener\" target=\"_blank\">microsoft.com\/windows\/windows-11-specifications<\/a>, <a href=\"https:\/\/docs.ollama.com\/gpu\" rel=\"noopener\" target=\"_blank\">docs.ollama.com\/gpu<\/a>, <a href=\"https:\/\/fastflowlm.com\" rel=\"noopener\" target=\"_blank\">fastflowlm.com<\/a><\/li>\n\n\n<li>Modellm\u00e9retek: <a href=\"https:\/\/huggingface.co\" rel=\"noopener\" target=\"_blank\">huggingface.co<\/a> (ggml-org, unsloth \u00e9s bartowski repozit\u00f3riumok)<\/li>\n\n\n<li>\u00c1rak: Apple Newsroom, NVIDIA Newsroom \u00e9s fejleszt\u0151i f\u00f3rumok, <a href=\"https:\/\/servethehome.com\" rel=\"noopener\" target=\"_blank\">servethehome.com<\/a>, <a href=\"https:\/\/storagereview.com\" rel=\"noopener\" target=\"_blank\">storagereview.com<\/a>, <a href=\"https:\/\/notebookcheck.net\" rel=\"noopener\" target=\"_blank\">notebookcheck.net<\/a>, <a href=\"https:\/\/tomshardware.com\" rel=\"noopener\" target=\"_blank\">tomshardware.com<\/a><\/li>\n\n<\/ol>\n\n\n\n<div class=\"kk-star-ratings kksr-auto kksr-align-center kksr-valign-bottom\"\n    data-payload='{&quot;align&quot;:&quot;center&quot;,&quot;id&quot;:&quot;29372&quot;,&quot;slug&quot;:&quot;default&quot;,&quot;valign&quot;:&quot;bottom&quot;,&quot;ignore&quot;:&quot;&quot;,&quot;reference&quot;:&quot;auto&quot;,&quot;class&quot;:&quot;&quot;,&quot;count&quot;:&quot;0&quot;,&quot;legendonly&quot;:&quot;&quot;,&quot;readonly&quot;:&quot;&quot;,&quot;score&quot;:&quot;0&quot;,&quot;starsonly&quot;:&quot;&quot;,&quot;best&quot;:&quot;5&quot;,&quot;gap&quot;:&quot;5&quot;,&quot;greet&quot;:&quot;Rate this post&quot;,&quot;legend&quot;:&quot;0\\\/5 - (0 votes)&quot;,&quot;size&quot;:&quot;24&quot;,&quot;title&quot;:&quot;Helyi AI-hardver 2026-ban: milyen sz\u00e1m\u00edt\u00f3g\u00e9p kell az LLM-ek otthoni futtat\u00e1s\u00e1hoz&quot;,&quot;width&quot;:&quot;0&quot;,&quot;_legend&quot;:&quot;{score}\\\/{best} - ({count} {votes})&quot;,&quot;font_factor&quot;:&quot;1.25&quot;}'>\n            \n<div class=\"kksr-stars\">\n    \n<div class=\"kksr-stars-inactive\">\n            <div class=\"kksr-star\" data-star=\"1\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" data-star=\"2\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" data-star=\"3\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" data-star=\"4\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" data-star=\"5\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n    <\/div>\n    \n<div class=\"kksr-stars-active\" style=\"width: 0px;\">\n            <div class=\"kksr-star\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n            <div class=\"kksr-star\" style=\"padding-right: 5px\">\n            \n\n<div class=\"kksr-icon\" style=\"width: 24px; height: 24px;\"><\/div>\n        <\/div>\n    <\/div>\n<\/div>\n                \n\n<div class=\"kksr-legend\" style=\"font-size: 19.2px;\">\n            <span class=\"kksr-muted\">Rate this post<\/span>\n    <\/div>\n    <\/div>\n","protected":false},"excerpt":{"rendered":"<p>A helyi AI-n\u00e1l nem a processzor d\u00f6nt, hanem a mem\u00f3ria m\u00e9rete \u00e9s s\u00e1vsz\u00e9less\u00e9ge. Elmagyar\u00e1zzuk, hogyan fut egy nyelvi modell, \u00e9s ellen\u0151rz\u00f6tt sz\u00e1mok alapj\u00e1n \u00f6sszehasonl\u00edtjuk a 2026-os hardvereket a videok\u00e1rty\u00e1kt\u00f3l a munka\u00e1llom\u00e1sokig.<\/p>\n","protected":false},"author":1,"featured_media":29371,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_kad_blocks_custom_css":"","_kad_blocks_head_custom_js":"","_kad_blocks_body_custom_js":"","_kad_blocks_footer_custom_js":"","_kad_post_transparent":"","_kad_post_title":"","_kad_post_layout":"","_kad_post_sidebar_id":"","_kad_post_content_style":"","_kad_post_vertical_padding":"","_kad_post_feature":"","_kad_post_feature_position":"","_kad_post_header":false,"_kad_post_footer":false,"_kad_post_classname":"","_wrc_toc":"on","_wrc_buttons_enabled":false,"_wrc_buttons":[],"hide-affiliate-disclaimer":"","_wrc_hide_banners":"","footnotes":""},"categories":[119,173,119],"tags":[],"class_list":["post-29372","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-mesterseges-intelligencia","category-guides"],"taxonomy_info":{"category":[{"value":119,"label":"Mesters\u00e9ges intelligencia"},{"value":173,"label":"Guides"},{"value":119,"label":"Mesters\u00e9ges intelligencia"}]},"featured_image_src_large":["https:\/\/www.websiteradar.net\/wp-content\/uploads\/2026\/10\/B3-18617-3-1024x585.png",1024,585,true],"author_info":{"display_name":"Tom\u00e1\u0161 Mahr\u00edk","author_link":"https:\/\/www.websiteradar.net\/hu\/author\/maho125\/"},"comment_info":0,"category_info":[{"term_id":119,"name":"Mesters\u00e9ges intelligencia","slug":"mesterseges-intelligencia","term_group":0,"term_taxonomy_id":119,"taxonomy":"category","description":"","parent":0,"count":3,"filter":"raw","cat_ID":119,"category_count":3,"category_description":"","cat_name":"Mesters\u00e9ges intelligencia","category_nicename":"mesterseges-intelligencia","category_parent":0},{"term_id":173,"name":"Guides","slug":"guides","term_group":0,"term_taxonomy_id":173,"taxonomy":"category","description":"","parent":0,"count":20,"filter":"raw","cat_ID":173,"category_count":20,"category_description":"","cat_name":"Guides","category_nicename":"guides","category_parent":0},{"term_id":119,"name":"Mesters\u00e9ges intelligencia","slug":"mesterseges-intelligencia","term_group":0,"term_taxonomy_id":119,"taxonomy":"category","description":"","parent":0,"count":3,"filter":"raw","cat_ID":119,"category_count":3,"category_description":"","cat_name":"Mesters\u00e9ges intelligencia","category_nicename":"mesterseges-intelligencia","category_parent":0}],"tag_info":false,"_links":{"self":[{"href":"https:\/\/www.websiteradar.net\/hu\/wp-json\/wp\/v2\/posts\/29372","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.websiteradar.net\/hu\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.websiteradar.net\/hu\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.websiteradar.net\/hu\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.websiteradar.net\/hu\/wp-json\/wp\/v2\/comments?post=29372"}],"version-history":[{"count":1,"href":"https:\/\/www.websiteradar.net\/hu\/wp-json\/wp\/v2\/posts\/29372\/revisions"}],"predecessor-version":[{"id":29395,"href":"https:\/\/www.websiteradar.net\/hu\/wp-json\/wp\/v2\/posts\/29372\/revisions\/29395"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.websiteradar.net\/hu\/wp-json\/wp\/v2\/media\/29371"}],"wp:attachment":[{"href":"https:\/\/www.websiteradar.net\/hu\/wp-json\/wp\/v2\/media?parent=29372"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.websiteradar.net\/hu\/wp-json\/wp\/v2\/categories?post=29372"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.websiteradar.net\/hu\/wp-json\/wp\/v2\/tags?post=29372"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}