{"id":478934,"date":"2026-05-07T11:47:45","date_gmt":"2026-05-07T11:47:45","guid":{"rendered":"https:\/\/savepearlharbor.com\/?p=478934"},"modified":"-0001-11-30T00:00:00","modified_gmt":"-0001-11-29T21:00:00","slug":"","status":"publish","type":"post","link":"https:\/\/savepearlharbor.com\/?p=478934","title":{"rendered":"\ufe0f Zyphra \u0432\u044b\u043a\u0430\u0442\u0438\u043b\u0430 ZAYA1-8B \u2014 \u043c\u0430\u043b\u0435\u043d\u044c\u043a\u0443\u044e MoE-\u043c\u043e\u0434\u0435\u043b\u044c, \u043a\u043e\u0442\u043e\u0440\u0430\u044f \u0432\u044b\u0433\u043b\u044f\u0434\u0438\u0442 \u0441\u043b\u0438\u0448\u043a\u043e\u043c \u0431\u043e\u0434\u0440\u043e \u0434\u043b\u044f \u0441\u0432\u043e\u0435\u0433\u043e \u0440\u0430\u0437\u043c\u0435\u0440\u0430"},"content":{"rendered":"<div xmlns=\"http:\/\/www.w3.org\/1999\/xhtml\">\n<p>\u0423 \u043c\u043e\u0434\u0435\u043b\u0438 \u043c\u0435\u043d\u044c\u0448\u0435 1 \u043c\u043b\u0440\u0434 \u0430\u043a\u0442\u0438\u0432\u043d\u044b\u0445 \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u043e\u0432, \u043d\u043e Zyphra \u0437\u0430\u044f\u0432\u043b\u044f\u0435\u0442, \u0447\u0442\u043e \u043e\u043d\u0430 \u043a\u043e\u043d\u043a\u0443\u0440\u0438\u0440\u0443\u0435\u0442 \u0441 \u043a\u0443\u0434\u0430 \u0431\u043e\u043b\u0435\u0435 \u043a\u0440\u0443\u043f\u043d\u044b\u043c\u0438 open-weight \u0438 proprietary-\u043c\u043e\u0434\u0435\u043b\u044f\u043c\u0438 \u043d\u0430 \u043c\u0430\u0442\u0435\u043c\u0430\u0442\u0438\u043a\u0435, \u043a\u043e\u0434\u0438\u043d\u0433\u0435 \u0438 reasoning-\u0431\u0435\u043d\u0447\u043c\u0430\u0440\u043a\u0430\u0445.<\/p>\n<p>ZAYA1-8B &#8212; \u044d\u0442\u043e \u0441\u0442\u0430\u0432\u043a\u0430 \u043d\u0430 \u0432\u0435\u0441\u044c \u0441\u0442\u0435\u043a \u0441\u0440\u0430\u0437\u0443:<\/p>\n<ul>\n<li>\n<p>MoE-\u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0430 \u0441 Compressed Convolutional Attention;<\/p>\n<\/li>\n<li>\n<p>\u043d\u043e\u0432\u044b\u0439 MLP-router \u0434\u043b\u044f \u0431\u043e\u043b\u0435\u0435 \u0441\u0442\u0430\u0431\u0438\u043b\u044c\u043d\u043e\u0433\u043e \u0432\u044b\u0431\u043e\u0440\u0430 \u044d\u043a\u0441\u043f\u0435\u0440\u0442\u043e\u0432;<\/p>\n<\/li>\n<li>\n<p>learned residual scaling \u0434\u043b\u044f \u043a\u043e\u043d\u0442\u0440\u043e\u043b\u044f \u0440\u043e\u0441\u0442\u0430 residual-\u043d\u043e\u0440\u043c;<\/p>\n<\/li>\n<li>\n<p>pretraining \u043d\u0430 AMD Instinct MI300x, \u0431\u0435\u0437 NVIDIA-\u0441\u0442\u0435\u043a\u0430;<\/p>\n<\/li>\n<li>\n<p>\u0431\u043e\u043b\u044c\u0448\u043e\u0439 post-training pipeline \u0441 SFT, reasoning warmup, RLVE-Gym, math\/code RL \u0438 RLHF\/RLAIF;<\/p>\n<\/li>\n<li>\n<p>test-time compute \u043c\u0435\u0442\u043e\u0434 Markovian RSA.<\/p>\n<\/li>\n<\/ul>\n<p>Markovian RSA &#8212; \u0441\u0430\u043c\u0430\u044f \u043b\u044e\u0431\u043e\u043f\u044b\u0442\u043d\u0430\u044f \u0447\u0430\u0441\u0442\u044c. \u041c\u043e\u0434\u0435\u043b\u044c \u0433\u0435\u043d\u0435\u0440\u0438\u0440\u0443\u0435\u0442 \u043d\u0435\u0441\u043a\u043e\u043b\u044c\u043a\u043e reasoning-\u0442\u0440\u0430\u0441\u0441 \u043f\u0430\u0440\u0430\u043b\u043b\u0435\u043b\u044c\u043d\u043e, \u0437\u0430\u0442\u0435\u043c \u0440\u0435\u043a\u0443\u0440\u0441\u0438\u0432\u043d\u043e \u0430\u0433\u0440\u0435\u0433\u0438\u0440\u0443\u0435\u0442 \u0438\u0445 \u0438 \u043f\u0440\u043e\u0434\u043e\u043b\u0436\u0430\u0435\u0442 \u0440\u0430\u0441\u0441\u0443\u0436\u0434\u0435\u043d\u0438\u0435 \u043a\u0443\u0441\u043a\u0430\u043c\u0438, \u043d\u0435 \u0440\u0430\u0437\u0434\u0443\u0432\u0430\u044f \u043a\u043e\u043d\u0442\u0435\u043a\u0441\u0442 \u0431\u0435\u0441\u043a\u043e\u043d\u0435\u0447\u043d\u043e. \u041f\u043e \u0437\u0430\u044f\u0432\u043b\u0435\u043d\u0438\u044e Zyphra, \u0438\u043c\u0435\u043d\u043d\u043e \u044d\u0442\u043e \u0440\u0435\u0437\u043a\u043e \u0431\u0443\u0441\u0442\u0438\u0442 \u0441\u043b\u043e\u0436\u043d\u044b\u0435 \u043c\u0430\u0442\u0435\u043c\u0430\u0442\u0438\u0447\u0435\u0441\u043a\u0438\u0435 \u0437\u0430\u0434\u0430\u0447\u0438.<\/p>\n<p>\u041d\u0430 HMMT\u201925 \u043e\u043d\u0438 \u043f\u043e\u043a\u0430\u0437\u044b\u0432\u0430\u044e\u0442 89.6 \u043f\u0440\u043e\u0442\u0438\u0432 88.3 \u0443 Claude 4.5 Sonnet \u0438 GPT-5-High. \u0410 \u043f\u0440\u0438 extra-high test-time compute \u0433\u043e\u0432\u043e\u0440\u044f\u0442, \u0447\u0442\u043e ZAYA1-8B \u043e\u0431\u0445\u043e\u0434\u0438\u0442 DeepSeek-V3.2 \u0438 GPT-OSS-120B High \u043d\u0430 APEX-shortlist.<\/p>\n<p>\u041c\u043e\u0434\u0435\u043b\u044c \u043f\u043e\u043b\u043d\u043e\u0441\u0442\u044c\u044e \u043e\u0431\u0443\u0447\u0430\u043b\u0438 \u043d\u0430 AMD-\u0438\u043d\u0444\u0440\u0430\u0441\u0442\u0440\u0443\u043a\u0442\u0443\u0440\u0435. \u0414\u043b\u044f \u0440\u044b\u043d\u043a\u0430 \u044d\u0442\u043e \u043f\u043e\u0447\u0442\u0438 \u043f\u043e\u043b\u0438\u0442\u0438\u0447\u0435\u0441\u043a\u043e\u0435 \u0437\u0430\u044f\u0432\u043b\u0435\u043d\u0438\u0435 &#8212; serious AI training \u0431\u043e\u043b\u044c\u0448\u0435 \u043d\u0435 \u043e\u0431\u044f\u0437\u0430\u043d \u0432\u044b\u0433\u043b\u044f\u0434\u0435\u0442\u044c \u043a\u0430\u043a \u201cNVIDIA or nothing\u201d.<\/p>\n<p>ZAYA1-8B \u0434\u043e\u0441\u0442\u0443\u043f\u043d\u0430 \u0432 Zyphra Cloud, \u0432\u0435\u0441\u0430 \u0432\u044b\u043b\u043e\u0436\u0435\u043d\u044b \u043d\u0430 Hugging Face, \u043b\u0438\u0446\u0435\u043d\u0437\u0438\u044f Apache-2.0. \u0422\u0435\u0441\u0442\u044b \u043e\u043f\u0443\u0431\u043b\u0438\u043a\u0443\u044e \u0443 \u0441\u0435\u0431\u044f \u0432 \u0442\u0433, \u0437\u0430\u0445\u043e\u0434\u0438\u0442\u0435: <a href=\"https:\/\/t.me\/rust_code\" rel=\"noopener noreferrer nofollow\">rust_code<\/a><\/p>\n<p><a href=\"https:\/\/www.zyphra.com\/post\/zaya1-8b\" rel=\"noopener noreferrer nofollow\">https:\/\/www.zyphra.com\/post\/zaya1-8b<\/a><\/p>\n<\/div>\n<p>\u0441\u0441\u044b\u043b\u043a\u0430 \u043d\u0430 \u043e\u0440\u0438\u0433\u0438\u043d\u0430\u043b \u0441\u0442\u0430\u0442\u044c\u0438 <a href=\"https:\/\/habr.com\/ru\/articles\/1032596\/\">https:\/\/habr.com\/ru\/articles\/1032596\/<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u0423 \u043c\u043e\u0434\u0435\u043b\u0438 \u043c\u0435\u043d\u044c\u0448\u0435 1 \u043c\u043b\u0440\u0434 \u0430\u043a\u0442\u0438\u0432\u043d\u044b\u0445 \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u043e\u0432, \u043d\u043e Zyphra \u0437\u0430\u044f\u0432\u043b\u044f\u0435\u0442, \u0447\u0442\u043e \u043e\u043d\u0430 \u043a\u043e\u043d\u043a\u0443\u0440\u0438\u0440\u0443\u0435\u0442 \u0441 \u043a\u0443\u0434\u0430 \u0431\u043e\u043b\u0435\u0435 \u043a\u0440\u0443\u043f\u043d\u044b\u043c\u0438 open-weight \u0438 proprietary-\u043c\u043e\u0434\u0435\u043b\u044f\u043c\u0438 \u043d\u0430 \u043c\u0430\u0442\u0435\u043c\u0430\u0442\u0438\u043a\u0435, \u043a\u043e\u0434\u0438\u043d\u0433\u0435 \u0438 reasoning-\u0431\u0435\u043d\u0447\u043c\u0430\u0440\u043a\u0430\u0445.ZAYA1-8B &#8212; \u044d\u0442\u043e \u0441\u0442\u0430\u0432\u043a\u0430 \u043d\u0430 \u0432\u0435\u0441\u044c \u0441\u0442\u0435\u043a \u0441\u0440\u0430\u0437\u0443:MoE-\u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0430 \u0441 Compressed Convolutional Attention;\u043d\u043e\u0432\u044b\u0439 MLP-router \u0434\u043b\u044f \u0431\u043e\u043b\u0435\u0435 \u0441\u0442\u0430\u0431\u0438\u043b\u044c\u043d\u043e\u0433\u043e \u0432\u044b\u0431\u043e\u0440\u0430 \u044d\u043a\u0441\u043f\u0435\u0440\u0442\u043e\u0432;learned residual scaling \u0434\u043b\u044f \u043a\u043e\u043d\u0442\u0440\u043e\u043b\u044f \u0440\u043e\u0441\u0442\u0430 residual-\u043d\u043e\u0440\u043c;pretraining \u043d\u0430 AMD Instinct MI300x, \u0431\u0435\u0437 NVIDIA-\u0441\u0442\u0435\u043a\u0430;\u0431\u043e\u043b\u044c\u0448\u043e\u0439 post-training pipeline \u0441 SFT, reasoning warmup, RLVE-Gym, math\/code RL \u0438 RLHF\/RLAIF;test-time compute \u043c\u0435\u0442\u043e\u0434 Markovian RSA.Markovian RSA &#8212; \u0441\u0430\u043c\u0430\u044f \u043b\u044e\u0431\u043e\u043f\u044b\u0442\u043d\u0430\u044f \u0447\u0430\u0441\u0442\u044c. \u041c\u043e\u0434\u0435\u043b\u044c \u0433\u0435\u043d\u0435\u0440\u0438\u0440\u0443\u0435\u0442 \u043d\u0435\u0441\u043a\u043e\u043b\u044c\u043a\u043e reasoning-\u0442\u0440\u0430\u0441\u0441 \u043f\u0430\u0440\u0430\u043b\u043b\u0435\u043b\u044c\u043d\u043e, \u0437\u0430\u0442\u0435\u043c \u0440\u0435\u043a\u0443\u0440\u0441\u0438\u0432\u043d\u043e \u0430\u0433\u0440\u0435\u0433\u0438\u0440\u0443\u0435\u0442 \u0438\u0445 \u0438 \u043f\u0440\u043e\u0434\u043e\u043b\u0436\u0430\u0435\u0442 \u0440\u0430\u0441\u0441\u0443\u0436\u0434\u0435\u043d\u0438\u0435 \u043a\u0443\u0441\u043a\u0430\u043c\u0438, \u043d\u0435 \u0440\u0430\u0437\u0434\u0443\u0432\u0430\u044f \u043a\u043e\u043d\u0442\u0435\u043a\u0441\u0442 \u0431\u0435\u0441\u043a\u043e\u043d\u0435\u0447\u043d\u043e. \u041f\u043e \u0437\u0430\u044f\u0432\u043b\u0435\u043d\u0438\u044e Zyphra, \u0438\u043c\u0435\u043d\u043d\u043e \u044d\u0442\u043e \u0440\u0435\u0437\u043a\u043e \u0431\u0443\u0441\u0442\u0438\u0442 \u0441\u043b\u043e\u0436\u043d\u044b\u0435 \u043c\u0430\u0442\u0435\u043c\u0430\u0442\u0438\u0447\u0435\u0441\u043a\u0438\u0435 \u0437\u0430\u0434\u0430\u0447\u0438.\u041d\u0430 HMMT\u201925 \u043e\u043d\u0438 \u043f\u043e\u043a\u0430\u0437\u044b\u0432\u0430\u044e\u0442 89.6 \u043f\u0440\u043e\u0442\u0438\u0432 88.3 \u0443 Claude 4.5 Sonnet \u0438 GPT-5-High. \u0410 \u043f\u0440\u0438 extra-high test-time compute \u0433\u043e\u0432\u043e\u0440\u044f\u0442, \u0447\u0442\u043e ZAYA1-8B \u043e\u0431\u0445\u043e\u0434\u0438\u0442 DeepSeek-V3.2 \u0438 GPT-OSS-120B High \u043d\u0430 APEX-shortlist.\u041c\u043e\u0434\u0435\u043b\u044c \u043f\u043e\u043b\u043d\u043e\u0441\u0442\u044c\u044e \u043e\u0431\u0443\u0447\u0430\u043b\u0438 \u043d\u0430 AMD-\u0438\u043d\u0444\u0440\u0430\u0441\u0442\u0440\u0443\u043a\u0442\u0443\u0440\u0435. \u0414\u043b\u044f \u0440\u044b\u043d\u043a\u0430 \u044d\u0442\u043e \u043f\u043e\u0447\u0442\u0438 \u043f\u043e\u043b\u0438\u0442\u0438\u0447\u0435\u0441\u043a\u043e\u0435 \u0437\u0430\u044f\u0432\u043b\u0435\u043d\u0438\u0435 &#8212; serious AI training \u0431\u043e\u043b\u044c\u0448\u0435 \u043d\u0435 \u043e\u0431\u044f\u0437\u0430\u043d \u0432\u044b\u0433\u043b\u044f\u0434\u0435\u0442\u044c \u043a\u0430\u043a \u201cNVIDIA or nothing\u201d.ZAYA1-8B \u0434\u043e\u0441\u0442\u0443\u043f\u043d\u0430 \u0432 Zyphra Cloud, \u0432\u0435\u0441\u0430 \u0432\u044b\u043b\u043e\u0436\u0435\u043d\u044b \u043d\u0430 Hugging Face, \u043b\u0438\u0446\u0435\u043d\u0437\u0438\u044f Apache-2.0. \u0422\u0435\u0441\u0442\u044b \u043e\u043f\u0443\u0431\u043b\u0438\u043a\u0443\u044e \u0443 \u0441\u0435\u0431\u044f \u0432 \u0442\u0433, \u0437\u0430\u0445\u043e\u0434\u0438\u0442\u0435: rust_codehttps:\/\/www.zyphra.com\/post\/zaya1-8b\u0441\u0441\u044b\u043b\u043a\u0430 \u043d\u0430 \u043e\u0440\u0438\u0433\u0438\u043d\u0430\u043b \u0441\u0442\u0430\u0442\u044c\u0438 https:\/\/habr.com\/ru\/articles\/1032596\/<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[],"tags":[],"class_list":["post-478934","post","type-post","status-publish","format-standard","hentry"],"_links":{"self":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/posts\/478934","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=478934"}],"version-history":[{"count":0,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/posts\/478934\/revisions"}],"wp:attachment":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=478934"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=478934"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=478934"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}