{"id":158,"date":"2026-06-17T10:23:02","date_gmt":"2026-06-17T08:23:02","guid":{"rendered":"https:\/\/www.martinpavlicek.cz\/blog\/?p=158"},"modified":"2026-06-17T10:50:50","modified_gmt":"2026-06-17T08:50:50","slug":"novy-model-z-glm-5-2","status":"publish","type":"post","link":"https:\/\/www.martinpavlicek.cz\/blog\/novy-model-z-glm-5-2\/","title":{"rendered":"Nov\u00fd model – Z GLM 5.2"},"content":{"rendered":"\n

Je\u0161t\u011b ned\u00e1vno jsme se bavili o tom, \u017ee \u010c\u00edna doh\u00e1n\u00ed z\u00e1padn\u00ed AI laborato\u0159e. Dnes mi to zn\u00ed skoro zastarale. Modely jako DeepSeek, Kimi nebo te\u010f GLM-5.2 <\/a>od Z.ai ukazuj\u00ed, \u017ee \u010d\u00ednsk\u00fd AI ekosyst\u00e9m u\u017e nehraje jen hru na \u201elevn\u011bj\u0161\u00ed alternativu\u201c. Za\u010d\u00edn\u00e1 tla\u010dit p\u0159esn\u011b tam, kde to v\u00fdvoj\u00e1\u0159e bol\u00ed nejv\u00edc: dlouh\u00fd kontext, pr\u00e1ce s cel\u00fdm projektem, agentn\u00ed coding a praktick\u00e9 nasazen\u00ed.<\/p>\n\n\n\n

GLM-5.2 je podle Z.ai vlajkov\u00fd model zam\u011b\u0159en\u00fd na takzvan\u00e9 long-horizon tasks, tedy \u00falohy, kde model nem\u00e1 jen odpov\u011bd\u011bt na jednu ot\u00e1zku, ale dr\u017eet dlouh\u00fd pl\u00e1n, ch\u00e1pat v\u011bt\u0161\u00ed codebase, respektovat architekturu a postupn\u011b upravovat projekt od po\u017eadavk\u016f a\u017e po nasaditeln\u00fd v\u00fdsledek. Ofici\u00e1ln\u00ed dokumentace uv\u00e1d\u00ed kontext 1M token\u016f a maxim\u00e1ln\u00ed v\u00fdstup 128K token\u016f, co\u017e je p\u0159esn\u011b ten typ parametr\u016f, kter\u00e9 na pap\u00ed\u0159e vypadaj\u00ed jako marketing, ale v re\u00e1ln\u00e9m v\u00fdvoji mohou zm\u011bnit zp\u016fsob pr\u00e1ce s AI asistenty.<\/p>\n\n\n\n

Podle m\u011b je tu d\u016fle\u017eit\u00e1 jedna v\u011bc: dlouh\u00fd kontext s\u00e1m o sob\u011b nesta\u010d\u00ed<\/strong>. U\u017e jsme vid\u011bli modely, kter\u00e9 se chlubily obrovsk\u00fdm kontextov\u00fdm oknem, ale v praxi se v dlouh\u00fdch vstupech ztr\u00e1cely. Z.ai proto u GLM-5.2 zd\u016fraz\u0148uje, \u017ee nejde jen o \u201e1M token\u016f\u201c, ale o pou\u017eiteln\u00fd 1M kontext pro coding agenty, refaktoring, v\u00fdzkumnou reprodukci, mobiln\u00ed debugging nebo pr\u00e1ci nap\u0159\u00ed\u010d cel\u00fdm projektem. Dokumentace p\u0159\u00edmo popisuje sc\u00e9n\u00e1\u0159e jako p\u0159evzet\u00ed cel\u00e9 codebase, dlouhodob\u00fd refaktoring, dodr\u017eov\u00e1n\u00ed t\u00fdmov\u00fdch engineering standard\u016f nebo debugging p\u0159es ADB a logcat.<\/p>\n\n\n\n

A to je za m\u011b mnohem zaj\u00edmav\u011bj\u0161\u00ed ne\u017e dal\u0161\u00ed tabulka benchmark\u016f.<\/p>\n\n\n\n

Nech\u00e1pejte m\u011b \u0161patn\u011b, benchmarky jsou d\u016fle\u017eit\u00e9. Z.ai u GLM-5.2 tvrd\u00ed, \u017ee model dosahuje velmi siln\u00fdch v\u00fdsledk\u016f v long-horizon coding benchmarc\u00edch, nap\u0159\u00edklad FrontierSWE, PostTrainBench a SWE-Marathon, a \u017ee je mezi open-source modely na \u0161pi\u010dce. U standardn\u00edch coding benchmark\u016f firma uv\u00e1d\u00ed zlep\u0161en\u00ed proti GLM-5.1, nap\u0159\u00edklad 81.0 vs. 62.0 na Terminal-Bench 2.1 a 62.1 vs. 58.4 na SWE-bench Pro. Jen\u017ee tady bych byl opatrn\u00fd. V\u00fdsledky od v\u00fdrobce jsou fajn start, ale skute\u010dn\u00e1 pravda se uk\u00e1\u017ee a\u017e ve chv\u00edli, kdy model pust\u00edte na vlastn\u00ed repozit\u00e1\u0159, vlastn\u00ed legacy k\u00f3d a vlastn\u00ed chaotick\u00e9 zad\u00e1n\u00ed od klienta.<\/p>\n\n\n\n

Co se mi na GLM-5.2 l\u00edb\u00ed nejv\u00edc, je jeho pragmatick\u00e9 zam\u011b\u0159en\u00ed. Neprod\u00e1v\u00e1 se prim\u00e1rn\u011b jako chatbot pro v\u0161echno. Tla\u010d\u00ed se do role modelu pro v\u00fdvoj\u00e1\u0159e a coding agenty. Podporuje streaming, function calling, structured output, context caching a MCP integrace, tedy p\u0159esn\u011b ty v\u011bci, kter\u00e9 pot\u0159ebujete, kdy\u017e chcete model zapojit do re\u00e1ln\u00e9ho workflow, ne si s n\u00edm jen pov\u00eddat v okn\u011b prohl\u00ed\u017ee\u010de.<\/p>\n\n\n\n

Druh\u00e1 zaj\u00edmav\u00e1 v\u011bc je integrace do existuj\u00edc\u00edch n\u00e1stroj\u016f. GLM-5.2 se d\u00e1 podle dostupn\u00fdch n\u00e1vod\u016f pou\u017e\u00edvat p\u0159es Anthropic-compatible nebo OpenAI-compatible endpointy, tak\u017ee ho lze napojit do n\u00e1stroj\u016f typu Claude Code, Cline nebo OpenClaw bez toho, aby v\u00fdvoj\u00e1\u0159 musel kompletn\u011b m\u011bnit svoje prost\u0159ed\u00ed. DataCamp nap\u0159\u00edklad popisuje variantu glm-5.2[1m]<\/code>, kde se explicitn\u011b zap\u00edn\u00e1 1M tokenov\u00fd kontext, a upozor\u0148uje na dva reasoning re\u017eimy: high<\/code> a max<\/code>.<\/p>\n\n\n\n

Za m\u011b je tohle mo\u017en\u00e1 nejv\u011bt\u0161\u00ed posun. AI modely u\u017e nesout\u011b\u017e\u00ed jen v tom, kdo nap\u00ed\u0161e hez\u010d\u00ed odpov\u011b\u010f. Sout\u011b\u017e\u00ed v tom, kdo se l\u00e9pe napoj\u00ed do v\u00fdvoj\u00e1\u0159sk\u00e9ho procesu. Kdo um\u00ed \u010d\u00edst projekt. Kdo vydr\u017e\u00ed u \u00fakolu. Kdo nezapomene omezen\u00ed z prvn\u00ed \u010d\u00e1sti zad\u00e1n\u00ed. Kdo si nevyrob\u00ed vlastn\u00ed architekturu jen proto, \u017ee mu to vypad\u00e1 elegantn\u011bji.<\/p>\n\n\n\n

A tady se podle m\u011b l\u00e1me budoucnost AI ve v\u00fdvoji softwaru.<\/p>\n\n\n\n

GLM-5.2 bych ur\u010dit\u011b netestoval na trivi\u00e1ln\u00edm promptu typu \u201enapi\u0161 mi REST API v Node.js\u201c. To dnes zvl\u00e1dne skoro ka\u017ed\u00fd lep\u0161\u00ed model. Testoval bych ho na n\u011b\u010dem nep\u0159\u00edjemn\u00e9m: star\u0161\u00ed monolit, nejednotn\u00e9 testy, dokumentace mimo realitu, n\u011bkolik modul\u016f s historick\u00fdmi kompromisy a zad\u00e1n\u00ed typu \u201erefaktoruj to, ale nic nerozbij\u201c. Tam se uk\u00e1\u017ee, jestli 1M kontext znamen\u00e1 re\u00e1lnou v\u00fdhodu, nebo jen del\u0161\u00ed pam\u011b\u0165 pro sebev\u011bdom\u00e9 chyby.<\/p>\n\n\n\n

M\u016fj osobn\u00ed odhad? GLM-5.2 nebude model, kter\u00fd okam\u017eit\u011b v\u0161ichni vym\u011bn\u00ed za sv\u00e9 sou\u010dasn\u00e9 n\u00e1stroje. Ale je to dal\u0161\u00ed sign\u00e1l, \u017ee \u010d\u00ednsk\u00e9 modely se velmi rychle posouvaj\u00ed z kategorie \u201ezaj\u00edmav\u00e9 open-source alternativy\u201c do kategorie \u201etohle mus\u00edme re\u00e1ln\u011b benchmarkovat ve firm\u011b\u201c. A pokud se potvrd\u00ed stabilita na dlouh\u00fdch coding \u00faloh\u00e1ch, m\u016f\u017ee b\u00fdt pro v\u00fdvoj\u00e1\u0159sk\u00e9 t\u00fdmy hodn\u011b nep\u0159\u00edjemn\u00fdm konkurentem zaveden\u00fdch z\u00e1padn\u00edch model\u016f.<\/p>\n\n\n\n

Ne proto, \u017ee by m\u011bl nejlep\u0161\u00ed marketing.<\/p>\n\n\n\n

Ale proto, \u017ee m\u00ed\u0159\u00ed p\u0159esn\u011b na praktickou bolest: velk\u00fd projekt, dlouh\u00fd \u00fakol, mnoho soubor\u016f, hodn\u011b kontextu a minimum prostoru pro halucinace<\/strong>.<\/p>\n\n\n\n

A to je podle m\u011b sm\u011br, kter\u00fdm se AI coding bude v roce 2026 l\u00e1mat. ne chat, ale agentn\u00ed pr\u00e1ce nad cel\u00fdm softwarem.<\/p>\n","protected":false},"excerpt":{"rendered":"

Je\u0161t\u011b ned\u00e1vno jsme se bavili o tom, \u017ee \u010c\u00edna doh\u00e1n\u00ed z\u00e1padn\u00ed AI laborato\u0159e. Dnes mi to zn\u00ed skoro zastarale. Modely jako DeepSeek, Kimi nebo te\u010f GLM-5.2 od Z.ai ukazuj\u00ed, \u017ee \u010d\u00ednsk\u00fd AI ekosyst\u00e9m u\u017e nehraje jen hru na \u201elevn\u011bj\u0161\u00ed alternativu\u201c. Za\u010d\u00edn\u00e1 tla\u010dit p\u0159esn\u011b tam, kde to v\u00fdvoj\u00e1\u0159e bol\u00ed nejv\u00edc: dlouh\u00fd kontext, pr\u00e1ce s cel\u00fdm projektem, […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[31,74,28,75],"class_list":["post-158","post","type-post","status-publish","format-standard","hentry","category-nezarazene","tag-cina","tag-glm","tag-model","tag-z"],"_links":{"self":[{"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/posts\/158","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/comments?post=158"}],"version-history":[{"count":2,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/posts\/158\/revisions"}],"predecessor-version":[{"id":160,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/posts\/158\/revisions\/160"}],"wp:attachment":[{"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/media?parent=158"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/categories?post=158"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/tags?post=158"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}