{"id":158043,"date":"2026-04-15T18:17:12","date_gmt":"2026-04-15T18:17:12","guid":{"rendered":"https:\/\/www.europesays.com\/ro\/158043\/"},"modified":"2026-04-15T18:17:12","modified_gmt":"2026-04-15T18:17:12","slug":"gemma4-pe-windows-11-on-arm-copilot-snapdragon-x-elite-via-ollama-care-e-performanta-reala","status":"publish","type":"post","link":"https:\/\/www.europesays.com\/ro\/158043\/","title":{"rendered":"Gemma4 pe Windows 11 on ARM Copilot (Snapdragon X Elite) via Ollama: care e performanta reala?"},"content":{"rendered":"<p><img data-od-unknown-tag=\"\" data-od-xpath=\"\/HTML\/BODY\/DIV[@id='td-outer-wrap']\/*[2][self::DIV]\/*[2][self::DIV]\/*[1][self::DIV]\/*[1][self::ARTICLE]\/*[1][self::DIV]\/*[1][self::DIV]\/*[3][self::DIV]\/*[1][self::DIV]\/*[2][self::DIV]\/*[2][self::DIV]\/*[5][self::DIV]\/*[3][self::DIV]\/*[1][self::FIGURE]\/*[1][self::IMG]\" fetchpriority=\"high\" decoding=\"async\" width=\"1024\" height=\"855\" src=\"https:\/\/www.europesays.com\/ro\/wp-content\/uploads\/2026\/04\/ollama_gemma4_1-1024x855.jpg\" alt=\"Am reu\u0219it s\u0103 pun noul LLM Gemma4 pe un Asus Vivobook S15 Copilot+ PC (Snapdragon X Elite) cu Windows 11 on ARM ca s\u0103 v\u0103d ce poate \u0219i rezultatul e altul dec\u00e2t cel a\u0219teptat.\" class=\"wp-image-114193\"  \/>Am reu\u0219it s\u0103 pun noul LLM Gemma4 pe un Asus Vivobook S15 Copilot PC (Snapdragon X Elite) cu Windows 11 on ARM ca s\u0103 v\u0103d ce poate \u0219i rezultatul e altul dec\u00e2t cel a\u0219teptat.<\/p>\n<p>Pemntru c\u0103 am v\u0103zut c\u0103 noul model LLM<a href=\"https:\/\/www.mobzine.ro\/2026\/04\/cel-mai-nou-model-ai-gemma-4-poate-fi-folosit-local-fara-conexiune-la-net-pe-iphone-si-android\/\" type=\"link\" id=\"https:\/\/www.mobzine.ro\/2026\/04\/cel-mai-nou-model-ai-gemma-4-poate-fi-folosit-local-fara-conexiune-la-net-pe-iphone-si-android\/\" target=\"_blank\" rel=\"noreferrer noopener nofollow\"> Gemma4 lansat de Google poate fi folosit local<\/a> (complet offline), ca s\u0103 beneficiezi de AI pe smartphone, am zis c\u0103 merit\u0103 s\u0103 v\u0103d cum va func\u021biona el pe un Asus Vivobook S15 Copilot PC cu Windows 11 on ARM \u0219i 32GB RAM, sper\u00e2nd c\u0103 pot pune \u00een valoare NPU-ul ridicat at\u00e2ta \u00een sl\u0103vi de c\u0103tre Microsoft. Din p\u0103cate nu merge (\u00eenc\u0103!) pe NPU, \u00eens\u0103 se descurc\u0103 rulat pe CPU.<\/p>\n<p>Dar poate platforma Snapdragon X Elite (lansat\u0103 acum 2 ani)<a href=\"https:\/\/www.mobzine.ro\/2026\/04\/comparatie-samsung-galaxy-s26-utra-vs-xiaomi-17-ultra-performanta\/\" type=\"link\" id=\"https:\/\/www.mobzine.ro\/2026\/04\/comparatie-samsung-galaxy-s26-utra-vs-xiaomi-17-ultra-performanta\/\" target=\"_blank\" rel=\"noreferrer noopener nofollow\"> s\u0103 ofere performan\u021be m\u0103car la egalitate cu cipsetul Snapdragom SM8850-AC Elite Gen 5 (3 nm)<\/a> disponibil pe cele mai noi flagship-uri de la Samsung sau Xiaomi? \u0218i rezultatul m-a surprins.<\/p>\n<p>Cum rulezi Gemma4 local pe Asus Vivobook S15 Copilot+ PC cu Windows on ARM?<\/p>\n<p>Prima chestie pe care a trebuit s\u0103 o fac este s\u0103 g\u0103sesc o solu\u021bie de a rula local Gemma4 pe Windows 11 on ARM. Dup\u0103 ce m-am documentat pu\u021bin am ajuns la concluzia c\u0103 nu merge prin AnythingLLM via NPU, nu merge nici prin Microsoft Foundry via VS Code, tot pe NPU, \u00eens\u0103 merge pe CPU <a href=\"https:\/\/ollama.com\/\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">dac\u0103 folosesc Ollam<\/a>a.<\/p>\n<p>M-am desumflat rapid c\u00e2nd am \u00een\u021beles c\u0103 nu am cum s\u0103 m\u0103 folosesc de NPU pe un sistem Windows 11 on ARM Copilot \u0219i nici m\u0103car pe GPU, singura consolare era c\u0103 speram c\u0103 o s\u0103 se mi\u0219te m\u0103car acceptabil pe CPU.<\/p>\n<p>Dai fuga <a href=\"https:\/\/ollama.com\/download\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">pe site la Ollama<\/a>, descarci executabilul (din p\u0103cate nu e optimizat pentru Windows 11 on ARM) \u0219i \u00eel instalezi. Apoi descarci ce model Gemma4 vrei.<\/p>\n<p>Am mers direct pe varianta Gemma4:e2b, cea care \u00een teorie \u00eemi permitea cea mai bun\u0103 experien\u021b\u0103 \u2013 Ollama ofer\u0103 acces direct la acest model, trebuie doar s\u0103 \u00eel selectezi din interfa\u021b\u0103 \u0219i s\u0103 a\u0219tep\u021bi s\u0103 se descarce \u0219i s\u0103 se instaleze. Pe urm\u0103 \u00eel po\u021bi folosi, fie via API, fie \u00een interfa\u021b\u0103, fie \u00een mod CLI.<\/p>\n<p><img loading=\"lazy\" data-od-unknown-tag=\"\" data-od-xpath=\"\/HTML\/BODY\/DIV[@id='td-outer-wrap']\/*[2][self::DIV]\/*[2][self::DIV]\/*[1][self::DIV]\/*[1][self::ARTICLE]\/*[1][self::DIV]\/*[1][self::DIV]\/*[3][self::DIV]\/*[1][self::DIV]\/*[2][self::DIV]\/*[2][self::DIV]\/*[5][self::DIV]\/*[3][self::DIV]\/*[9][self::FIGURE]\/*[1][self::IMG]\" decoding=\"async\" width=\"1024\" height=\"371\" src=\"https:\/\/www.europesays.com\/ro\/wp-content\/uploads\/2026\/04\/ollama_gemma4-1024x371.jpg\" alt=\"\" class=\"wp-image-114194\"  \/><\/p>\n<p>I-am dat drumul \u00een interfa\u021b\u0103 ca s\u0103 v\u0103d cum merge \u0219i \u2026 a fost decent. Pentru o conversa\u021bie cu AI s-a descurcat chiar bine, cu toate c\u0103 procesorul \u00ee\u0219i duce toate core-urile spre 100% aproape instant.<\/p>\n<p>Urm\u0103torul pas a fost s\u0103 v\u0103d dac\u0103 face fa\u021b\u0103 cu un OpenCode \u00een docker \u2013 o dorin\u021b\u0103 mult prea optimist\u0103 pentru c\u0103 dup\u0103 mai mult de 20 de minute de procesat, nu a putut s\u0103 \u00eemi dea nici m\u0103car un r\u0103spuns la o problem\u0103 simpl\u0103 de Hello World \u00een Python.<\/p>\n<p>Asta a fost momentul \u00een care am renun\u021bat \u0219i am zis c\u0103 poate nu ar fi r\u0103u s\u0103 \u00eei fac un benchmark \u0219i s\u0103 v\u0103d cam cum st\u0103 treaba.<\/p>\n<p>Gemma4 pe Windows on ARM cu Snapdragon X Elite<\/p>\n<p>\u0218i a venit momentul adev\u0103rului \u2026 16.21 tokens\/s, pe Gemma4:e2b. Pute\u021bi testa \u0219i voi folosind urm\u0103toarea comand\u0103 \u00een PowerShell:<\/p>\n<p>ollama run gemma4:e2b &#8211;verbose<\/p>\n<p>\u0218i apoi \u00eei da\u021bi ce prompt dori\u021bi.<\/p>\n<p><img loading=\"lazy\" data-od-unknown-tag=\"\" data-od-xpath=\"\/HTML\/BODY\/DIV[@id='td-outer-wrap']\/*[2][self::DIV]\/*[2][self::DIV]\/*[1][self::DIV]\/*[1][self::ARTICLE]\/*[1][self::DIV]\/*[1][self::DIV]\/*[3][self::DIV]\/*[1][self::DIV]\/*[2][self::DIV]\/*[2][self::DIV]\/*[5][self::DIV]\/*[3][self::DIV]\/*[17][self::FIGURE]\/*[1][self::IMG]\" decoding=\"async\" width=\"627\" height=\"244\" src=\"https:\/\/www.europesays.com\/ro\/wp-content\/uploads\/2026\/04\/Screenshot-2026-04-10-202841.png\" alt=\"\" class=\"wp-image-114195\"  \/><\/p>\n<p>Recunosc c\u0103 m\u0103 a\u0219teptam la mai mult, m\u0103car la o valoare dubl\u0103. 16.21 \u2013 19.33 tokens pe secund\u0103 e o valoare mic\u0103, s\u0103 zicem suficient\u0103 pentru opera\u021bii simple de chat care s\u0103 implice generare de text, analiz\u0103 de text etc. \u00cens\u0103 nu pentru sarcini complexe \u0219i sub  nicio form\u0103 cu o alt\u0103 aplica\u021bie care s\u0103 consume putere de procesare \u00een fundal pentru c\u0103 posibil s\u0103 ajunge\u0219i s\u0103 nu mai pute\u021bi folosi PC-ul.<\/p>\n<p><img data-od-unknown-tag=\"\" data-od-xpath=\"\/HTML\/BODY\/DIV[@id='td-outer-wrap']\/*[2][self::DIV]\/*[2][self::DIV]\/*[1][self::DIV]\/*[1][self::ARTICLE]\/*[1][self::DIV]\/*[1][self::DIV]\/*[3][self::DIV]\/*[1][self::DIV]\/*[2][self::DIV]\/*[2][self::DIV]\/*[5][self::DIV]\/*[3][self::DIV]\/*[19][self::FIGURE]\/*[1][self::IMG]\" loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"559\" src=\"https:\/\/www.europesays.com\/ro\/wp-content\/uploads\/2026\/04\/ollama_gemma4_3-1024x559.jpg\" alt=\"\" class=\"wp-image-114201\"  \/><\/p>\n<p>Cu alte cuvinte, Gemma4:e2b merge local f\u0103r\u0103 preten\u021bii. Mi-am adus aminte c\u0103 acum 25 de ani, c\u00e2nd m\u0103 jucam pe PC-uri cu AMD\/Intel \u0219i \u00eencepusem s\u0103 prind gust de 3D cu pl\u0103cile 3Dfx \u0219i Banshee, m\u0103 bucuram dac\u0103 puteam s\u0103 joc ceva (F1 de exemplu) undeva la 15-20 fps. Cam acela\u0219i sentiment l-am avut \u0219i acum cu Gemma4 local.<\/p>\n<p>Cum merge Gemma4 pe un flagship Android?<\/p>\n<p>\u0218i dac\u0103 tot am pornit pe drumul \u0103sta \u00een care m-am jucat cu Gemma4 \u0219i \u00eenc\u0103 testez Xiaomi 17 Ultra \u0219i Samsung Galaxy S26 Ultra, mi-a fost clar c\u0103 trebuie s\u0103 m\u0103sor \u0219i cum se mi\u0219c\u0103 AI-ul cu Gemma4 pe ele.<\/p>\n<p>Folosind Google AI Edge Gallery am desc\u0103rcat modelele Gemma4 e2b \u0219i e4b pe Xiaomi 17 Ultra \u0219i le-am pus la un test care a spus urm\u0103toarele:<\/p>\n<tr>\n<td>Model<\/td>\n<td>CPU (tokens\/s)<\/td>\n<td>GPU (tokens\/s_<\/td>\n<\/tr>\n<tr>\n<td>Gemma4:e4b<\/td>\n<td>6.17<\/td>\n<td>16.16<\/td>\n<\/tr>\n<tr>\n<td>Gemma4:e2b<\/td>\n<td>16.10<\/td>\n<td>39.32<\/td>\n<\/tr>\n<p><img data-od-unknown-tag=\"\" data-od-xpath=\"\/HTML\/BODY\/DIV[@id='td-outer-wrap']\/*[2][self::DIV]\/*[2][self::DIV]\/*[1][self::DIV]\/*[1][self::ARTICLE]\/*[1][self::DIV]\/*[1][self::DIV]\/*[3][self::DIV]\/*[1][self::DIV]\/*[2][self::DIV]\/*[2][self::DIV]\/*[5][self::DIV]\/*[3][self::DIV]\/*[25][self::FIGURE]\/*[1][self::IMG]\" loading=\"lazy\" decoding=\"async\" width=\"981\" height=\"1024\" src=\"https:\/\/www.europesays.com\/ro\/wp-content\/uploads\/2026\/04\/Screenshot-2026-04-15-204016-981x1024.png\" alt=\"\" class=\"wp-image-114198\"  \/><\/p>\n<p>Pe e4b am mers \u00een test cu 1500 tokens pentru prefill, iar pentru e2b doar cu 512 (ca s\u0103 se termine mai repede testul):<\/p>\n<p><img data-od-unknown-tag=\"\" data-od-xpath=\"\/HTML\/BODY\/DIV[@id='td-outer-wrap']\/*[2][self::DIV]\/*[2][self::DIV]\/*[1][self::DIV]\/*[1][self::ARTICLE]\/*[1][self::DIV]\/*[1][self::DIV]\/*[3][self::DIV]\/*[1][self::DIV]\/*[2][self::DIV]\/*[2][self::DIV]\/*[5][self::DIV]\/*[3][self::DIV]\/*[27][self::FIGURE]\/*[1][self::FIGURE]\/*[1][self::IMG]\" loading=\"lazy\" decoding=\"async\" width=\"471\" height=\"1024\" data-id=\"114199\" src=\"https:\/\/www.europesays.com\/ro\/wp-content\/uploads\/2026\/04\/3-471x1024.jpg\" alt=\"\" class=\"wp-image-114199\"  \/><\/p>\n<p><img data-od-unknown-tag=\"\" data-od-xpath=\"\/HTML\/BODY\/DIV[@id='td-outer-wrap']\/*[2][self::DIV]\/*[2][self::DIV]\/*[1][self::DIV]\/*[1][self::ARTICLE]\/*[1][self::DIV]\/*[1][self::DIV]\/*[3][self::DIV]\/*[1][self::DIV]\/*[2][self::DIV]\/*[2][self::DIV]\/*[5][self::DIV]\/*[3][self::DIV]\/*[27][self::FIGURE]\/*[2][self::FIGURE]\/*[1][self::IMG]\" loading=\"lazy\" decoding=\"async\" width=\"500\" height=\"1087\" data-id=\"114200\" src=\"https:\/\/www.europesays.com\/ro\/wp-content\/uploads\/2026\/04\/4-471x1024.jpg\" alt=\"\" class=\"wp-image-114200\"  \/><\/p>\n<p>\u00cen traducere, situa\u021bia st\u0103 \u00een felul urm\u0103tor: dac\u0103 alege\u021bi modeul cu 2 miliarde de parametrii, pe GPU, Xiaomi 17 Ultra merge excep\u021bional, cu aproape 40 tokens pe secund\u0103 (dac\u0103 \u0219i pe Windows on ARM Gemma4 ar fi mers la fel, era super bine), iar pe CPU e la 16 tokens. Cam acelea\u0219i date le-am ob\u021binut cu AI Edge Gallery \u0219i pe iPhone 16 Pro, unde am avut 34.38 tokens\/s \u0103e GPU \u0219i 32.72 tokens\/s \u00een CPU.<\/p>\n<p>Pe modelul cu 4 miliarde de parametrii, performan\u021ba e la sub jum\u0103tate \u0219i pe CPU \u0219i pe GPU, dar \u00eenc\u0103 folosibil pentru sarcini simple. Practic pot s\u0103 spun c\u0103 Gemma4 e4b merge pe un flagship Android curent (\u00een GPU) cu aceia\u0219i vitez\u0103 cu care func\u021bioneaz\u0103 Gemma4 e2b pe un Copilot PC cu Snapdragon X Elite, dar pe CPU.<\/p>\n<p>De re\u021binut c\u0103 \u0219i pe CPU \u0219i pe GPU, smartphone-urile Android se incing semnificativ \u0219i bateria scade vertiginos. Iar pe PC-ul Copilot, CPU-ul se duce rapid \u00een 100% pentru aproape toate cele 12 core-uri \u0219i are un impact major \u00een func\u021bionare, orice alte aplica\u021bii de pe l\u00e2ng\u0103 Ollama se vor mi\u0219ca \u00eengrozitor de \u00eencet.<\/p>\n<p>Concluzie: merge, dar cam at\u00e2t<\/p>\n<p>\u00cen concluzie, dac\u0103 vre\u021bi s\u0103 v\u0103 juca\u021bi local cu Gemma4 pe un PC Copilot cu Windows 11 on ARM trebuie s\u0103 \u0219ti\u021bi c\u0103 nu merge nici pe NPU, nici pe GPU (cel pu\u021bin acum) \u0219i \u00eel pute\u021bi folosi doar pe CPU. Unde consum\u0103 CPU-ul \u0219i aduce valoare limitat\u0103.<\/p>\n<p>\u0218i nicidecum nu poate face fa\u021b\u0103 unor sarcini complexe, \u00een special cele programare \u00eempreun\u0103 cu OpenCode.<\/p>\n<p>Pe de alt\u0103 parte, dac\u0103 ave\u021bi un flagship modern (c\u0103 e Android sau iOS), v\u0103 pute\u021bi juca \u00een voie cu el pe GPU, unde cele aproape 40 tokens\/secund\u0103 v\u0103 ofer\u0103 o experien\u021b\u0103 foarte bun\u0103. E \u0219i normal, Google a optimizat Gemma4 e2b pentru a putea fi folosit local, pe smartphone. Doar c\u0103 suge bateria la greu.<\/p>\n","protected":false},"excerpt":{"rendered":"Am reu\u0219it s\u0103 pun noul LLM Gemma4 pe un Asus Vivobook S15 Copilot PC (Snapdragon X Elite) cu&hellip;\n","protected":false},"author":2,"featured_media":158044,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[16],"tags":[43581,83,9669,4994,42503,16723,46055,46056,5337,1725,41,40,38,39,163,46057,141,124,2453,169],"class_list":{"0":"post-158043","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-tehnologie","8":"tag-17-ultra","9":"tag-apple","10":"tag-arm","11":"tag-galaxy-s26-ultra","12":"tag-llm","13":"tag-local","14":"tag-ollama","15":"tag-opencode","16":"tag-performanta","17":"tag-programare","18":"tag-ro","19":"tag-romana","20":"tag-romania","21":"tag-romanian","22":"tag-samsung","23":"tag-snapdragon-x-elite","24":"tag-technology","25":"tag-tehnologie","26":"tag-windows-11","27":"tag-xiaomi"},"share_on_mastodon":{"url":"https:\/\/pubeurope.com\/@ro\/116410098887510773","error":""},"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ro\/wp-json\/wp\/v2\/posts\/158043","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ro\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ro\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ro\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ro\/wp-json\/wp\/v2\/comments?post=158043"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ro\/wp-json\/wp\/v2\/posts\/158043\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ro\/wp-json\/wp\/v2\/media\/158044"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ro\/wp-json\/wp\/v2\/media?parent=158043"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ro\/wp-json\/wp\/v2\/categories?post=158043"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ro\/wp-json\/wp\/v2\/tags?post=158043"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}