{"id":139,"date":"2026-06-01T18:42:13","date_gmt":"2026-06-01T16:42:13","guid":{"rendered":"https:\/\/www.martinpavlicek.cz\/blog\/?p=139"},"modified":"2026-06-17T11:03:15","modified_gmt":"2026-06-17T09:03:15","slug":"novy-model-minimax-m3","status":"publish","type":"post","link":"https:\/\/www.martinpavlicek.cz\/blog\/novy-model-minimax-m3\/","title":{"rendered":"Nov\u00fd model MiniMax M3"},"content":{"rendered":"\n

Kdy\u017e jsem se d\u00edval na nov\u00fd MiniMax M3<\/strong><\/a>, prvn\u00ed v\u011bc, kter\u00e1 m\u011b zaujala, nebylo jen klasick\u00e9 marketingov\u00e9 \u201em\u00e1me lep\u0161\u00ed benchmarky\u201c. To dnes tvrd\u00ed ka\u017ed\u00fd. Zaj\u00edmav\u011bj\u0161\u00ed je pro m\u011b kombinace t\u0159\u00ed v\u011bc\u00ed: agentn\u00ed pr\u00e1ce, programov\u00e1n\u00ed a extr\u00e9mn\u011b dlouh\u00fd kontext<\/strong>. A pr\u00e1v\u011b tahle kombinace m\u016f\u017ee b\u00fdt pro v\u00fdvoj\u00e1\u0159e a firmy d\u016fle\u017eit\u011bj\u0161\u00ed ne\u017e samotn\u00e9 \u010d\u00edslo v n\u011bjak\u00e9m leaderboardu.<\/p>\n\n\n\n

MiniMax M3 byl ofici\u00e1ln\u011b p\u0159edstaven 1. \u010dervna 2026. Firma ho popisuje jako model zam\u011b\u0159en\u00fd na coding, agentn\u00ed \u00falohy, multimodalitu a pr\u00e1ci s kontextem a\u017e 1 milion token\u016f<\/strong>, p\u0159i\u010dem\u017e u API uv\u00e1d\u00ed garantovan\u00e9 minimum 512 tis\u00edc token\u016f<\/strong>.<\/p>\n\n\n\n

Pro\u010d m\u011b M3 zaujal<\/h2>\n\n\n\n

Podle m\u011b se u LLM model\u016f pomalu l\u00e1me jedna \u00e9ra. Nesta\u010d\u00ed, aby model dob\u0159e odpov\u011bd\u011bl na ot\u00e1zku. St\u00e1le v\u00edc p\u016fjde o to, jestli dok\u00e1\u017ee pracovat dlouhodob\u011b<\/strong>, dr\u017eet kontext, pou\u017e\u00edvat n\u00e1stroje, opravovat vlastn\u00ed chyby a postupn\u011b se posouvat k v\u00fdsledku.<\/p>\n\n\n\n

MiniMax p\u0159esn\u011b t\u00edmto sm\u011brem M3 prezentuje. Mluv\u00ed o autonomn\u00edm rozkladu \u00faloh, vol\u00e1n\u00ed n\u00e1stroj\u016f, v\u00edcekrokov\u00e9m uva\u017eov\u00e1n\u00ed a vyu\u017eit\u00ed v AI coding asistentech nebo automatizovan\u00fdch workflow.<\/p>\n\n\n\n

A tady si mysl\u00edm, \u017ee je z\u00e1sadn\u00ed rozd\u00edl mezi \u201emodelem na odpov\u011bdi\u201c a \u201emodelem na pr\u00e1ci\u201c. Chatbot odpov\u00ed. Agent mus\u00ed n\u011bco ud\u011blat<\/strong>.<\/p>\n\n\n\n

1M kontext: ne jen v\u011bt\u0161\u00ed \u010d\u00edslo<\/h2>\n\n\n\n

Kontextov\u00e9 okno o velikosti 1M token\u016f zn\u00ed na prvn\u00ed pohled jako technick\u00e1 chlouba. Jen\u017ee v praxi m\u016f\u017ee znamenat n\u011bco velmi konkr\u00e9tn\u00edho: modelu d\u00e1te v\u011bt\u0161\u00ed \u010d\u00e1st repozit\u00e1\u0159e, dokumentace, log\u016f, po\u017eadavk\u016f, star\u0161\u00edch rozhodnut\u00ed a on s t\u00edm m\u016f\u017ee pracovat najednou.<\/p>\n\n\n\n

To je obrovsk\u00fd rozd\u00edl oproti situaci, kdy se model neust\u00e1le pt\u00e1 na chyb\u011bj\u00edc\u00ed soubory nebo zapom\u00edn\u00e1, co se \u0159e\u0161ilo p\u0159ed chv\u00edl\u00ed. MiniMax tvrd\u00ed, \u017ee za t\u00edm stoj\u00ed jejich architektura MiniMax Sparse Attention<\/strong>, kter\u00e1 m\u00e1 \u0161k\u00e1lovat dlouh\u00fd kontext efektivn\u011bji ne\u017e klasick\u00e1 pln\u00e1 attention. Firma uv\u00e1d\u00ed, \u017ee p\u0159i d\u00e9lce 1 milion token\u016f m\u00e1 M3 per-token compute pouze na \u00farovni 1\/20 p\u0159edchoz\u00ed generace a \u017ee dos\u00e1hli v\u00edce ne\u017e 9\u00d7 zrychlen\u00ed v prefillingu a v\u00edce ne\u017e 15\u00d7 ve f\u00e1zi decodingu.<\/p>\n\n\n\n

Tady bych ale byl opatrn\u00fd. Jsou to zat\u00edm hlavn\u011b \u010d\u00edsla od v\u00fdrobce. Beru je jako zaj\u00edmav\u00fd sign\u00e1l, ne jako definitivn\u00ed pravdu.<\/p>\n\n\n\n

Coding: oblast, kde se l\u00e1me d\u016fv\u011bra<\/h2>\n\n\n\n

U programov\u00e1n\u00ed m\u011b zaj\u00edm\u00e1 hlavn\u011b jedna ot\u00e1zka: dok\u00e1\u017ee model dlouhodob\u011b pracovat na re\u00e1ln\u00e9m probl\u00e9mu, nebo jen hezky vygeneruje kus k\u00f3du?<\/p>\n\n\n\n

MiniMax uv\u00e1d\u00ed v\u00fdsledky nap\u0159\u00edklad SWE-Bench Pro 59,0 %<\/strong>, Terminal-Bench 2.1 66,0 %<\/strong>, SWE-fficiency 34,8 %<\/strong>, KernelBench Hard 28,8 %<\/strong> a MCP Atlas 74,2 %<\/strong>.<\/p>\n\n\n\n

To jsou zaj\u00edmav\u00e1 \u010d\u00edsla, ale pro m\u011b je d\u016fle\u017eit\u011bj\u0161\u00ed jin\u00e1 \u010d\u00e1st prezentace: MiniMax zd\u016fraz\u0148uje simulaci re\u00e1ln\u00e9 spolupr\u00e1ce s v\u00fdvoj\u00e1\u0159em. Tedy ne jen jednor\u00e1zov\u00fd prompt, ale pr\u016fb\u011b\u017en\u00e9 up\u0159es\u0148ov\u00e1n\u00ed po\u017eadavk\u016f, iterace, korekce a p\u0159ep\u00edn\u00e1n\u00ed mezi \u00falohami.<\/p>\n\n\n\n

A to je p\u0159esn\u011b to, kde dne\u0161n\u00ed coding asistenti \u010dasto selh\u00e1vaj\u00ed. Um\u00ed navrhnout \u0159e\u0161en\u00ed, ale kdy\u017e se projekt za\u010dne v\u011btvit, naraz\u00ed na limity pam\u011bti, pl\u00e1nov\u00e1n\u00ed a sebekontroly.<\/p>\n\n\n\n

Agentn\u00ed schopnosti: nejsiln\u011bj\u0161\u00ed p\u0159\u00edslib i nejv\u011bt\u0161\u00ed riziko<\/h2>\n\n\n\n

MiniMax ukazuje n\u011bkolik p\u016fsobiv\u00fdch experiment\u016f. M3 m\u011bl nap\u0159\u00edklad autonomn\u011b reprodukovat pr\u00e1ci ocen\u011bnou na ICLR 2025, b\u011b\u017eet t\u00e9m\u011b\u0159 12 hodin, vytvo\u0159it 18 commit\u016f a 23 experiment\u00e1ln\u00edch graf\u016f. V jin\u00e9m testu optimalizoval CUDA kernel p\u0159ibli\u017en\u011b 24 hodin, provedl 147 benchmarkov\u00fdch submit\u016f a 1 959 tool call\u016f, p\u0159i\u010dem\u017e m\u011bl dos\u00e1hnout 9,4\u00d7 zrychlen\u00ed.<\/p>\n\n\n\n

Tohle je podle m\u011b fascinuj\u00edc\u00ed. Ne proto, \u017ee bych automaticky v\u011b\u0159il ka\u017ed\u00e9mu \u010d\u00edslu, ale proto\u017ee to ukazuje sm\u011br: modely se neposouvaj\u00ed jen k lep\u0161\u00edm odpov\u011bd\u00edm, ale k dlouhodob\u00e9 autonomn\u00ed pr\u00e1ci<\/strong>.<\/p>\n\n\n\n

A z\u00e1rove\u0148 je to varov\u00e1n\u00ed. \u010c\u00edm d\u00e9le agent b\u011b\u017e\u00ed, t\u00edm d\u016fle\u017eit\u011bj\u0161\u00ed je audit, logov\u00e1n\u00ed, sandbox, limity opr\u00e1vn\u011bn\u00ed a mo\u017enost \u010dlov\u011bka zas\u00e1hnout. Agent, kter\u00fd ud\u011bl\u00e1 1 959 tool call\u016f, m\u016f\u017ee b\u00fdt produktivn\u00ed. Ale bez kontroly m\u016f\u017ee b\u00fdt tak\u00e9 drah\u00fd, chaotick\u00fd nebo nebezpe\u010dn\u00fd.<\/p>\n\n\n\n

Multimodalita od za\u010d\u00e1tku, ne jako p\u0159\u00edv\u011bsek<\/h2>\n\n\n\n

MiniMax tvrd\u00ed, \u017ee M3 je nativn\u011b multimod\u00e1ln\u00ed model, tedy \u017ee multimodalita nen\u00ed dodate\u010dn\u011b p\u0159ilepen\u00e1 vrstva, ale sou\u010d\u00e1st tr\u00e9ninku od za\u010d\u00e1tku. Model m\u00e1 podporovat obrazov\u00fd a video vstup a podle popisu MiniMaxu m\u016f\u017ee fungovat i v re\u017eimu ovl\u00e1d\u00e1n\u00ed po\u010d\u00edta\u010de.<\/p>\n\n\n\n

Tohle je podle m\u011b d\u016fle\u017eit\u00e9 hlavn\u011b pro firemn\u00ed workflow. Skute\u010dn\u00e1 pr\u00e1ce toti\u017e nen\u00ed jen text. M\u00e1me PDFka, screenshoty, grafy, tabulky, UI aplikac\u00ed, videa, logy, Excel soubory a intern\u00ed n\u00e1stroje. Model, kter\u00fd toto dok\u00e1\u017ee spojit, je mnohem bl\u00ed\u017e re\u00e1ln\u00e9mu digit\u00e1ln\u00edmu pracovn\u00edkovi ne\u017e klasick\u00fd textov\u00fd chatbot.<\/p>\n\n\n\n

Open-weight model: pokud to MiniMax dot\u00e1hne, bude to zaj\u00edmav\u00e9<\/h2>\n\n\n\n

MiniMax M3 prezentuje jako prvn\u00ed open-weight model, kter\u00fd spojuje frontier coding, milionov\u00fd kontext a nativn\u00ed multimodalitu. Na str\u00e1nce z\u00e1rove\u0148 uv\u00e1d\u00ed, \u017ee model m\u00e1 b\u00fdt pln\u011b open-source na Hugging Face a GitHubu, s podporou priv\u00e1tn\u00edho nasazen\u00ed a fine-tuningu.<\/p>\n\n\n\n

Pro m\u011b je tohle mo\u017en\u00e1 nejzaj\u00edmav\u011bj\u0161\u00ed \u010d\u00e1st. Pokud bude model skute\u010dn\u011b prakticky pou\u017eiteln\u00fd i mimo uzav\u0159en\u00e9 API, otev\u00edr\u00e1 to dve\u0159e firm\u00e1m, kter\u00e9 necht\u011bj\u00ed pos\u00edlat citliv\u00fd k\u00f3d nebo dokumenty do ciz\u00edho cloudu.<\/p>\n\n\n\n

Samoz\u0159ejm\u011b bude z\u00e1le\u017eet na licenci, re\u00e1ln\u00fdch hardwarov\u00fdch n\u00e1roc\u00edch, kvalit\u011b inference a dostupnosti vah. \u201eOpen-weight\u201c je\u0161t\u011b automaticky neznamen\u00e1 jednoduch\u00e9 lok\u00e1ln\u00ed nasazen\u00ed.<\/p>\n\n\n\n

M\u016fj n\u00e1zor<\/h2>\n\n\n\n

MiniMax M3 beru jako jeden z model\u016f, kter\u00e9 ukazuj\u00ed, kam se AI bude d\u00e1l posouvat: od chatbot\u016f k dlouhodob\u011b pracuj\u00edc\u00edm agent\u016fm<\/strong>. Nejde jen o to, \u017ee model odpov\u00ed l\u00e9pe. Jde o to, \u017ee zvl\u00e1dne dr\u017eet dlouh\u00fd kontext, pou\u017e\u00edvat n\u00e1stroje, pracovat s multimod\u00e1ln\u00edmi vstupy a postupn\u011b \u0159e\u0161it komplexn\u00ed \u00fakol.<\/p>\n\n\n\n

Z\u00e1rove\u0148 bych z\u016fstal st\u0159\u00edzliv\u00fd. Hodn\u011b \u00fadaj\u016f poch\u00e1z\u00ed p\u0159\u00edmo od MiniMaxu, tak\u017ee bych cht\u011bl vid\u011bt nez\u00e1visl\u00e9 testy, re\u00e1ln\u00e9 zku\u0161enosti v\u00fdvoj\u00e1\u0159\u016f a hlavn\u011b chov\u00e1n\u00ed modelu v b\u011b\u017en\u00fdch projektech, ne jen v pe\u010dliv\u011b p\u0159ipraven\u00fdch demonstra\u010dn\u00edch \u00faloh\u00e1ch.<\/p>\n\n\n\n

Ale sm\u011br je jasn\u00fd. Podle m\u011b u\u017e se nept\u00e1me jen: \u201eKter\u00fd model je chyt\u0159ej\u0161\u00ed?\u201c<\/strong>
Za\u010d\u00edn\u00e1me se pt\u00e1t: \u201eKter\u00e9mu modelu m\u016f\u017eu sv\u011b\u0159it kus skute\u010dn\u00e9 pr\u00e1ce?\u201c<\/strong><\/p>\n\n\n\n

A MiniMax M3 se evidentn\u011b sna\u017e\u00ed b\u00fdt odpov\u011bd\u00ed pr\u00e1v\u011b na tu druhou ot\u00e1zku.<\/p>\n","protected":false},"excerpt":{"rendered":"

Kdy\u017e jsem se d\u00edval na nov\u00fd MiniMax M3, prvn\u00ed v\u011bc, kter\u00e1 m\u011b zaujala, nebylo jen klasick\u00e9 marketingov\u00e9 \u201em\u00e1me lep\u0161\u00ed benchmarky\u201c. To dnes tvrd\u00ed ka\u017ed\u00fd. Zaj\u00edmav\u011bj\u0161\u00ed je pro m\u011b kombinace t\u0159\u00ed v\u011bc\u00ed: agentn\u00ed pr\u00e1ce, programov\u00e1n\u00ed a extr\u00e9mn\u011b dlouh\u00fd kontext. A pr\u00e1v\u011b tahle kombinace m\u016f\u017ee b\u00fdt pro v\u00fdvoj\u00e1\u0159e a firmy d\u016fle\u017eit\u011bj\u0161\u00ed ne\u017e samotn\u00e9 \u010d\u00edslo v n\u011bjak\u00e9m leaderboardu. […]<\/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":[31,65,28],"class_list":["post-139","post","type-post","status-publish","format-standard","hentry","category-nezarazene","tag-cina","tag-minimax","tag-model"],"_links":{"self":[{"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/posts\/139","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=139"}],"version-history":[{"count":2,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/posts\/139\/revisions"}],"predecessor-version":[{"id":164,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/posts\/139\/revisions\/164"}],"wp:attachment":[{"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/media?parent=139"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/categories?post=139"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/tags?post=139"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}