{"id":857,"date":"2025-02-04T16:45:13","date_gmt":"2025-02-04T16:45:13","guid":{"rendered":"https:\/\/janusai.pro\/?p=857"},"modified":"2025-02-04T16:45:15","modified_gmt":"2025-02-04T16:45:15","slug":"in-depth-analysis-what-innovations-does-deepseeks-newly-released-janus-pro-have","status":"publish","type":"post","link":"https:\/\/janusai.pro\/tr\/in-depth-analysis-what-innovations-does-deepseeks-newly-released-janus-pro-have\/","title":{"rendered":"Derinlemesine analiz: DeepSeek'in yeni piyasaya s\u00fcrd\u00fc\u011f\u00fc Janus-Pro hangi yeniliklere sahip?"},"content":{"rendered":"<div style=\"margin-top: 0px; margin-bottom: 0px;\" class=\"sharethis-inline-share-buttons\" ><\/div>\n<p>DeepSeek web sitesini g\u00fcncelledi.<\/p>\n\n\n\n<p>Y\u0131lba\u015f\u0131 gecesinin erken saatlerinde DeepSeek aniden GitHub'da Janus proje alan\u0131n\u0131n Janus-Pro modeli ve teknik raporunun kayna\u011f\u0131n\u0131 a\u00e7t\u0131\u011f\u0131n\u0131 duyurdu.<\/p>\n\n\n\n<p>\u00d6ncelikle, birka\u00e7 kilit noktan\u0131n alt\u0131n\u0131 \u00e7izelim:<\/p>\n\n\n\n<ol start=\"1\" class=\"wp-block-list\">\n<li>Bu <a href=\"https:\/\/huggingface.co\/deepseek-ai\/Janus-Pro-7B\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">Janus-Pro modeli<\/a> bu kez piyasaya s\u00fcr\u00fclen \u00e7ok modlu bir modeldir.<strong> \u00e7ok modlu anlama ve g\u00f6r\u00fcnt\u00fc olu\u015fturma g\u00f6revlerini ayn\u0131 anda ger\u00e7ekle\u015ftirebilir. Toplam iki parametre versiyonuna sahiptir, <a href=\"https:\/\/huggingface.co\/deepseek-ai\/Janus-Pro-7B\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">Janus-Pro-1B ve Janus-Pro-7B<\/a>.<\/strong><\/li>\n\n\n\n<li>Janus-Pro'nin temel yenili\u011fi, a\u015fa\u011f\u0131dakileri birbirinden ay\u0131rmakt\u0131r<strong> iki farkl\u0131 g\u00f6rev olan \u00e7ok modlu anlama ve \u00fcretme. Bu, bu iki g\u00f6revin ayn\u0131 modelde verimli bir \u015fekilde tamamlanmas\u0131na olanak tan\u0131r<\/strong>.<\/li>\n\n\n\n<li>Janus-Pro, DeepSeek taraf\u0131ndan ge\u00e7en Ekim ay\u0131nda piyasaya s\u00fcr\u00fclen Janus model mimarisiyle uyumludur, ancak o zaman Janus'un fazla hacmi yoktu. G\u00f6rme alan\u0131nda bir algoritma uzman\u0131 olan Dr. Charles bize \u00f6nceki Janus'un \"ortalama\" oldu\u011funu ve \"DeepSeek'in dil modeli kadar iyi olmad\u0131\u011f\u0131n\u0131\" s\u00f6yledi.<\/li>\n<\/ol>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1870\" height=\"1804\" src=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/720ad345-4902-46e0-9185-bc4f887ed302.png\" alt=\"\" class=\"wp-image-859\" srcset=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/720ad345-4902-46e0-9185-bc4f887ed302.png 1870w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/720ad345-4902-46e0-9185-bc4f887ed302-300x289.png 300w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/720ad345-4902-46e0-9185-bc4f887ed302-1024x988.png 1024w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/720ad345-4902-46e0-9185-bc4f887ed302-768x741.png 768w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/720ad345-4902-46e0-9185-bc4f887ed302-1536x1482.png 1536w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/720ad345-4902-46e0-9185-bc4f887ed302-12x12.png 12w\" sizes=\"auto, (max-width: 1870px) 100vw, 1870px\" \/><\/figure>\n\n\n\n<div id=\"ez-toc-container\" class=\"ez-toc-v2_0_82_2 counter-hierarchy ez-toc-counter ez-toc-grey ez-toc-container-direction\">\n<div class=\"ez-toc-title-container\">\n<p class=\"ez-toc-title\" style=\"cursor:inherit\">\u0130\u00e7indekiler<\/p>\n<span class=\"ez-toc-title-toggle\"><a href=\"#\" class=\"ez-toc-pull-right ez-toc-btn ez-toc-btn-xs ez-toc-btn-default ez-toc-toggle\" aria-label=\"\u0130\u00e7erik Tablosunu De\u011fi\u015ftir\"><span class=\"ez-toc-js-icon-con\"><span class=\"\"><span class=\"eztoc-hide\" style=\"display:none;\">Ge\u00e7i\u015f<\/span><span class=\"ez-toc-icon-toggle-span\"><svg style=\"fill: #999;color:#999\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" class=\"list-377408\" width=\"20px\" height=\"20px\" viewbox=\"0 0 24 24\" fill=\"none\"><path d=\"M6 6H4v2h2V6zm14 0H8v2h12V6zM4 11h2v2H4v-2zm16 0H8v2h12v-2zM4 16h2v2H4v-2zm16 0H8v2h12v-2z\" fill=\"currentColor\"><\/path><\/svg><svg style=\"fill: #999;color:#999\" class=\"arrow-unsorted-368013\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" width=\"10px\" height=\"10px\" viewbox=\"0 0 24 24\" version=\"1.2\" baseprofile=\"tiny\"><path d=\"M18.2 9.3l-6.2-6.3-6.2 6.3c-.2.2-.3.4-.3.7s.1.5.3.7c.2.2.4.3.7.3h11c.3 0 .5-.1.7-.3.2-.2.3-.5.3-.7s-.1-.5-.3-.7zM5.8 14.7l6.2 6.3 6.2-6.3c.2-.2.3-.5.3-.7s-.1-.5-.3-.7c-.2-.2-.4-.3-.7-.3h-11c-.3 0-.5.1-.7.3-.2.2-.3.5-.3.7s.1.5.3.7z\"\/><\/svg><\/span><\/span><\/span><\/a><\/span><\/div>\n<nav><ul class='ez-toc-list ez-toc-list-level-1' ><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-1\" href=\"https:\/\/janusai.pro\/tr\/in-depth-analysis-what-innovations-does-deepseeks-newly-released-janus-pro-have\/#It_is_intended_to_solve_the_industrys_difficult_problem_balancing_multimodal_understanding_and_image_generation\" >Sekt\u00f6r\u00fcn zor sorununu \u00e7\u00f6zmeyi ama\u00e7lamaktad\u0131r: \u00e7ok modlu anlama ve g\u00f6r\u00fcnt\u00fc olu\u015fturmay\u0131 dengelemek<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-2\" href=\"https:\/\/janusai.pro\/tr\/in-depth-analysis-what-innovations-does-deepseeks-newly-released-janus-pro-have\/#Januss_decoupled_architecture_and_Janus-Pros_optimized_training_strategy\" >Janus'un ayr\u0131\u015ft\u0131r\u0131lm\u0131\u015f mimarisi ve Janus-Pro'nin optimize edilmi\u015f e\u011fitim stratejisi<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-3\" href=\"https:\/\/janusai.pro\/tr\/in-depth-analysis-what-innovations-does-deepseeks-newly-released-janus-pro-have\/#First_lets_look_at_the_parameters\" >\u0130lk olarak parametrelere bakal\u0131m.<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-4\" href=\"https:\/\/janusai.pro\/tr\/in-depth-analysis-what-innovations-does-deepseeks-newly-released-janus-pro-have\/#There_is_also_the_training_strategy\" >Bir de e\u011fitim stratejisi var.<\/a><ul class='ez-toc-list-level-4' ><li class='ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-5\" href=\"https:\/\/janusai.pro\/tr\/in-depth-analysis-what-innovations-does-deepseeks-newly-released-janus-pro-have\/#Stage_I_%E2%80%93_Longer_training_time\" >A\u015fama I - Daha uzun e\u011fitim s\u00fcresi<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-6\" href=\"https:\/\/janusai.pro\/tr\/in-depth-analysis-what-innovations-does-deepseeks-newly-released-janus-pro-have\/#Stage_II_%E2%80%93_Removing_ImageNet_data_and_adding_multi-modal_data\" >A\u015fama II - ImageNet verilerinin kald\u0131r\u0131lmas\u0131 ve \u00e7ok modlu verilerin eklenmesi<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-7\" href=\"https:\/\/janusai.pro\/tr\/in-depth-analysis-what-innovations-does-deepseeks-newly-released-janus-pro-have\/#Stage_III_%E2%80%93_Optimizing_the_data_ratio\" >A\u015fama III - Veri oran\u0131n\u0131n optimize edilmesi<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-8\" href=\"https:\/\/janusai.pro\/tr\/in-depth-analysis-what-innovations-does-deepseeks-newly-released-janus-pro-have\/#Lets_look_at_the_training_data\" >E\u011fitim verilerine bakal\u0131m.<\/a><ul class='ez-toc-list-level-4' ><li class='ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-9\" href=\"https:\/\/janusai.pro\/tr\/in-depth-analysis-what-innovations-does-deepseeks-newly-released-janus-pro-have\/#Expansion_of_multimodal_understanding_data\" >\u00c7ok modlu anlama verilerinin geni\u015fletilmesi:<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-10\" href=\"https:\/\/janusai.pro\/tr\/in-depth-analysis-what-innovations-does-deepseeks-newly-released-janus-pro-have\/#Expansion_of_visual_generation_data\" >G\u00f6rsel \u00fcretim verilerinin geni\u015fletilmesi:<\/a><\/li><\/ul><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-11\" href=\"https:\/\/janusai.pro\/tr\/in-depth-analysis-what-innovations-does-deepseeks-newly-released-janus-pro-have\/#The_continuation_of_an_efficiency_revolution\" >Bir verimlilik devriminin devam\u0131 m\u0131?<\/a><\/li><\/ul><\/nav><\/div>\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"It_is_intended_to_solve_the_industrys_difficult_problem_balancing_multimodal_understanding_and_image_generation\"><\/span><strong>Sekt\u00f6r\u00fcn zor sorununu \u00e7\u00f6zmeyi ama\u00e7lamaktad\u0131r: \u00e7ok modlu anlama ve g\u00f6r\u00fcnt\u00fc olu\u015fturmay\u0131 dengelemek<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p>DeepSeek'in resmi tan\u0131t\u0131m\u0131na g\u00f6re, <a href=\"https:\/\/huggingface.co\/deepseek-ai\/Janus-Pro-7B\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">Janus-Pro <\/a>sadece resimleri anlamak, resimlerdeki metni \u00e7\u0131karmak ve anlamakla kalmaz, ayn\u0131 zamanda resimler de \u00fcretebilir.<\/p>\n\n\n\n<p>Teknik raporda, ayn\u0131 t\u00fcr ve b\u00fcy\u00fckl\u00fckteki di\u011fer modellerle kar\u015f\u0131la\u015ft\u0131r\u0131ld\u0131\u011f\u0131nda, Janus-Pro-7B'nin GenEval ve DPG-Bench test setlerindeki puanlar\u0131n\u0131n<strong> SD3-Medium ve DALL-E 3 gibi di\u011fer modellerden daha y\u00fcksektir.<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"1067\" src=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/a30e3dd3-b95c-4745-a4f7-8c5ace36af17.png\" alt=\"\" class=\"wp-image-862\" srcset=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/a30e3dd3-b95c-4745-a4f7-8c5ace36af17.png 1080w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/a30e3dd3-b95c-4745-a4f7-8c5ace36af17-300x296.png 300w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/a30e3dd3-b95c-4745-a4f7-8c5ace36af17-1024x1012.png 1024w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/a30e3dd3-b95c-4745-a4f7-8c5ace36af17-768x759.png 768w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/a30e3dd3-b95c-4745-a4f7-8c5ace36af17-12x12.png 12w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<p>Yetkili ayr\u0131ca \u00f6rnekler de veriyor \ud83d\udc47:<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"1295\" src=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/a063e5ec-bc07-4129-8ded-0ab90001bbfe.png\" alt=\"\" class=\"wp-image-866\" srcset=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/a063e5ec-bc07-4129-8ded-0ab90001bbfe.png 1080w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/a063e5ec-bc07-4129-8ded-0ab90001bbfe-250x300.png 250w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/a063e5ec-bc07-4129-8ded-0ab90001bbfe-854x1024.png 854w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/a063e5ec-bc07-4129-8ded-0ab90001bbfe-768x921.png 768w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/a063e5ec-bc07-4129-8ded-0ab90001bbfe-10x12.png 10w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<p>X'te yeni \u00f6zellikleri deneyen \u00e7ok say\u0131da netizen de var.<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"1429\" src=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/c5235f43-349a-47a2-ae9f-dc5491d88d61.png\" alt=\"\" class=\"wp-image-867\" srcset=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/c5235f43-349a-47a2-ae9f-dc5491d88d61.png 1080w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/c5235f43-349a-47a2-ae9f-dc5491d88d61-227x300.png 227w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/c5235f43-349a-47a2-ae9f-dc5491d88d61-774x1024.png 774w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/c5235f43-349a-47a2-ae9f-dc5491d88d61-768x1016.png 768w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/c5235f43-349a-47a2-ae9f-dc5491d88d61-9x12.png 9w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"1616\" src=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/61c76adf-1b00-4b38-8504-a47410160d3b.png\" alt=\"\" class=\"wp-image-869\" srcset=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/61c76adf-1b00-4b38-8504-a47410160d3b.png 1080w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/61c76adf-1b00-4b38-8504-a47410160d3b-200x300.png 200w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/61c76adf-1b00-4b38-8504-a47410160d3b-684x1024.png 684w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/61c76adf-1b00-4b38-8504-a47410160d3b-768x1149.png 768w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/61c76adf-1b00-4b38-8504-a47410160d3b-1027x1536.png 1027w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/61c76adf-1b00-4b38-8504-a47410160d3b-8x12.png 8w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<p>Ancak ara s\u0131ra \u00e7\u00f6kmeler de oluyor.<\/p>\n\n\n\n<p>Teknik dok\u00fcmanlara ba\u015fvurarak <a href=\"https:\/\/www.deepseek.com\/\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">DeepSeek<\/a>Janus Pro'nin \u00fc\u00e7 ay \u00f6nce piyasaya s\u00fcr\u00fclen Janus'a dayal\u0131 bir optimizasyon oldu\u011funu tespit ettik.<\/p>\n\n\n\n<p>Bu model serisinin temel yenili\u011fi \u015fudur<strong> G\u00f6rsel anlama g\u00f6revlerini g\u00f6rsel \u00fcretim g\u00f6revlerinden ay\u0131r\u0131n, b\u00f6ylece iki g\u00f6revin etkileri dengelenebilir.<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"538\" src=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/0fc71a9c-e23b-4ae9-976c-a4820124628e.png\" alt=\"\" class=\"wp-image-861\" srcset=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/0fc71a9c-e23b-4ae9-976c-a4820124628e.png 1080w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/0fc71a9c-e23b-4ae9-976c-a4820124628e-300x149.png 300w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/0fc71a9c-e23b-4ae9-976c-a4820124628e-1024x510.png 1024w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/0fc71a9c-e23b-4ae9-976c-a4820124628e-768x383.png 768w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/0fc71a9c-e23b-4ae9-976c-a4820124628e-18x9.png 18w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<p>Bir modelin ayn\u0131 anda \u00e7ok modlu anlama ve \u00fcretim ger\u00e7ekle\u015ftirmesi nadir de\u011fildir. Bu test setindeki D-DiT ve TokenFlow-XL bu yetene\u011fe sahiptir.<\/p>\n\n\n\n<p>Bununla birlikte, Janus'un karakteristik \u00f6zelli\u011fi \u015fudur<strong> \u0130\u015flemeyi ayr\u0131\u015ft\u0131rarak, \u00e7ok modlu anlama ve \u00fcretme yapabilen bir model, iki g\u00f6revin etkinli\u011fini dengeler.<\/strong><\/p>\n\n\n\n<p><strong>\u0130ki g\u00f6revin etkinli\u011fini dengelemek sekt\u00f6rde zor bir sorundur.<\/strong> \u00d6nceden, \u00e7ok modlu anlama ve \u00fcretimi m\u00fcmk\u00fcn oldu\u011funca uygulamak i\u00e7in ayn\u0131 kodlay\u0131c\u0131y\u0131 kullanma d\u00fc\u015f\u00fcncesi vard\u0131.<\/p>\n\n\n\n<p>Bu yakla\u015f\u0131m\u0131n avantajlar\u0131 basit bir mimari, gereksiz da\u011f\u0131t\u0131m olmamas\u0131 ve metin modelleriyle (metin olu\u015fturma ve metin anlama i\u00e7in ayn\u0131 y\u00f6ntemleri kullanan) uyumlu olmas\u0131d\u0131r. Bir ba\u015fka arg\u00fcman da, \u00e7oklu yeteneklerin bu \u015fekilde bir araya getirilmesinin belirli bir derecede ortaya \u00e7\u0131kmaya yol a\u00e7abilece\u011fidir.<\/p>\n\n\n\n<p>Bununla birlikte, asl\u0131nda, \u00fcretim ve anlamay\u0131 birle\u015ftirdikten sonra, iki g\u00f6rev \u00e7at\u0131\u015facakt\u0131r - g\u00f6r\u00fcnt\u00fc anlama, modelin y\u00fcksek boyutlarda soyutlama yapmas\u0131n\u0131 ve makroskopik olana e\u011filimli olan resmin temel anlam\u0131n\u0131 \u00e7\u0131karmas\u0131n\u0131 gerektirir. \u00d6te yandan g\u00f6r\u00fcnt\u00fc \u00fcretimi, piksel d\u00fczeyinde yerel ayr\u0131nt\u0131lar\u0131n ifade edilmesi ve \u00fcretilmesine odaklan\u0131r.<\/p>\n\n\n\n<p>Sekt\u00f6r\u00fcn ola\u011fan uygulamas\u0131, g\u00f6r\u00fcnt\u00fc olu\u015fturma yeteneklerine \u00f6ncelik vermektir. Bu da \u00e7ok modlu modellerle sonu\u00e7lan\u0131r<strong> daha y\u00fcksek kaliteli g\u00f6r\u00fcnt\u00fcler \u00fcretebilir, ancak g\u00f6r\u00fcnt\u00fc anlama sonu\u00e7lar\u0131 genellikle vasatt\u0131r.<\/strong><\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"Januss_decoupled_architecture_and_Janus-Pros_optimized_training_strategy\"><\/span><strong>Janus'un ayr\u0131\u015ft\u0131r\u0131lm\u0131\u015f mimarisi ve Janus-Pro'nin optimize edilmi\u015f e\u011fitim stratejisi<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p>Janus'un ayr\u0131\u015ft\u0131r\u0131lm\u0131\u015f mimarisi, modelin anlama ve \u00fcretme g\u00f6revlerini kendi ba\u015f\u0131na dengelemesini sa\u011flar.<\/p>\n\n\n\n<p>Resmi teknik raporda yer alan sonu\u00e7lara g\u00f6re, ister \u00e7ok modlu anlama ister g\u00f6r\u00fcnt\u00fc olu\u015fturma olsun, Janus-Pro-7B birden fazla test setinde iyi performans g\u00f6steriyor.<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"976\" src=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/f280e5d6-7150-40d3-bf81-7e5e5b780129.png\" alt=\"\" class=\"wp-image-863\" srcset=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/f280e5d6-7150-40d3-bf81-7e5e5b780129.png 1080w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/f280e5d6-7150-40d3-bf81-7e5e5b780129-300x271.png 300w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/f280e5d6-7150-40d3-bf81-7e5e5b780129-1024x925.png 1024w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/f280e5d6-7150-40d3-bf81-7e5e5b780129-768x694.png 768w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/f280e5d6-7150-40d3-bf81-7e5e5b780129-13x12.png 13w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<p><strong>\u00c7ok modlu anlay\u0131\u015f i\u00e7in,<\/strong> Janus-Pro-7B yedi de\u011ferlendirme veri setinin d\u00f6rd\u00fcnde birinci, kalan \u00fc\u00e7\u00fcnde ise en \u00fcst s\u0131radaki modelin biraz gerisinde kalarak ikinci olmu\u015ftur.<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"1062\" src=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/782c767f-e8d3-4d5c-9e2d-638c102f7dff.png\" alt=\"\" class=\"wp-image-865\" srcset=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/782c767f-e8d3-4d5c-9e2d-638c102f7dff.png 1080w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/782c767f-e8d3-4d5c-9e2d-638c102f7dff-300x295.png 300w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/782c767f-e8d3-4d5c-9e2d-638c102f7dff-1024x1007.png 1024w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/782c767f-e8d3-4d5c-9e2d-638c102f7dff-768x755.png 768w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/782c767f-e8d3-4d5c-9e2d-638c102f7dff-12x12.png 12w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<p><strong>G\u00f6r\u00fcnt\u00fc olu\u015fturma i\u00e7in,<\/strong> Janus-Pro-7B, hem GenEval hem de DPG-Bench de\u011ferlendirme veri k\u00fcmelerinde genel puanlamada birincili\u011fi elde etti.<\/p>\n\n\n\n<p>Bu \u00e7oklu g\u00f6rev etkisi temel olarak Janus serisinin farkl\u0131 g\u00f6revler i\u00e7in iki g\u00f6rsel kodlay\u0131c\u0131 kullanmas\u0131ndan kaynaklanmaktad\u0131r:<\/p>\n\n\n\n<ol start=\"1\" class=\"wp-block-list\">\n<li><strong>Kodlay\u0131c\u0131y\u0131 anlama:<\/strong> g\u00f6r\u00fcnt\u00fc anlama g\u00f6revleri (g\u00f6r\u00fcnt\u00fc sorular\u0131 ve cevaplar\u0131, g\u00f6rsel s\u0131n\u0131fland\u0131rma vb.) i\u00e7in g\u00f6r\u00fcnt\u00fclerdeki anlamsal \u00f6zellikleri \u00e7\u0131karmak i\u00e7in kullan\u0131l\u0131r.<\/li>\n\n\n\n<li><strong>\u00dcretken kodlay\u0131c\u0131:<\/strong> Metin-g\u00f6r\u00fcnt\u00fc olu\u015fturma g\u00f6revleri i\u00e7in g\u00f6r\u00fcnt\u00fcleri ayr\u0131k bir g\u00f6sterime d\u00f6n\u00fc\u015ft\u00fcr\u00fcr (\u00f6rne\u011fin, bir VQ kodlay\u0131c\u0131 kullanarak).<\/li>\n<\/ol>\n\n\n\n<p>Bu mimari ile,<strong> Model, her bir kodlay\u0131c\u0131n\u0131n performans\u0131n\u0131 ba\u011f\u0131ms\u0131z olarak optimize edebilir, b\u00f6ylece \u00e7ok modlu anlama ve \u00fcretme g\u00f6revlerinin her biri en iyi performansa ula\u015fabilir.<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"565\" src=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/81defdea-eadb-4fc7-8395-4c365c44d502.png\" alt=\"\" class=\"wp-image-860\" srcset=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/81defdea-eadb-4fc7-8395-4c365c44d502.png 1080w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/81defdea-eadb-4fc7-8395-4c365c44d502-300x157.png 300w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/81defdea-eadb-4fc7-8395-4c365c44d502-1024x536.png 1024w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/81defdea-eadb-4fc7-8395-4c365c44d502-768x402.png 768w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/81defdea-eadb-4fc7-8395-4c365c44d502-18x9.png 18w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<p><strong>Bu ayr\u0131\u015ft\u0131r\u0131lm\u0131\u015f mimari Janus-Pro ve Janus i\u00e7in ortakt\u0131r. Peki, Janus-Pro ge\u00e7ti\u011fimiz birka\u00e7 ay i\u00e7inde hangi yinelemeleri yapt\u0131?<\/strong><\/p>\n\n\n\n<p>De\u011ferlendirme setinin sonu\u00e7lar\u0131ndan g\u00f6r\u00fclebilece\u011fi gibi, Janus-Pro-1B'nin mevcut s\u00fcr\u00fcm\u00fc, \u00f6nceki Janus'a k\u0131yasla farkl\u0131 de\u011ferlendirme setlerinin puanlar\u0131nda yakla\u015f\u0131k 10% ila 20%'lik bir iyile\u015fmeye sahiptir. Janus-Pro-7B, parametre say\u0131s\u0131n\u0131 geni\u015flettikten sonra Janus'a k\u0131yasla yakla\u015f\u0131k 45%'lik en y\u00fcksek iyile\u015fmeye sahiptir.<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"185\" src=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/746d1d6f-9a07-4922-9b59-717614ef0738.png\" alt=\"\" class=\"wp-image-858\" srcset=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/746d1d6f-9a07-4922-9b59-717614ef0738.png 1080w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/746d1d6f-9a07-4922-9b59-717614ef0738-300x51.png 300w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/746d1d6f-9a07-4922-9b59-717614ef0738-1024x175.png 1024w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/746d1d6f-9a07-4922-9b59-717614ef0738-768x132.png 768w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/746d1d6f-9a07-4922-9b59-717614ef0738-18x3.png 18w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"430\" src=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/9361ef86-87ce-4f84-9cdd-71ef95a9f4b1.png\" alt=\"\" class=\"wp-image-864\" srcset=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/9361ef86-87ce-4f84-9cdd-71ef95a9f4b1.png 1080w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/9361ef86-87ce-4f84-9cdd-71ef95a9f4b1-300x119.png 300w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/9361ef86-87ce-4f84-9cdd-71ef95a9f4b1-1024x408.png 1024w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/9361ef86-87ce-4f84-9cdd-71ef95a9f4b1-768x306.png 768w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/9361ef86-87ce-4f84-9cdd-71ef95a9f4b1-18x7.png 18w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<p>E\u011fitim detaylar\u0131 a\u00e7\u0131s\u0131ndan, teknik raporda, Janus-Pro'nin mevcut s\u00fcr\u00fcm\u00fcn\u00fcn, \u00f6nceki Janus modeline k\u0131yasla, temel ayr\u0131\u015ft\u0131r\u0131lm\u0131\u015f mimari tasar\u0131m\u0131n\u0131 korudu\u011fu ve ek olarak a\u015fa\u011f\u0131dakileri yineledi\u011fi belirtilmektedir<strong> parametre boyutu, e\u011fitim stratejisi ve e\u011fitim verileri.<\/strong><\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"First_lets_look_at_the_parameters\"><\/span><strong>\u0130lk olarak, parametrelere bakal\u0131m<\/strong>.<span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>Janus'un ilk s\u00fcr\u00fcm\u00fcnde sadece 1.3B parametre vard\u0131 ve Pro'nun mevcut s\u00fcr\u00fcm\u00fc 1B ve 7B parametreli modeller i\u00e7eriyor.<\/p>\n\n\n\n<p>Bu iki boyut Janus mimarisinin \u00f6l\u00e7eklenebilirli\u011fini yans\u0131tmaktad\u0131r. En hafif olan 1B modeli, WebGPU kullan\u0131larak taray\u0131c\u0131da \u00e7al\u0131\u015ft\u0131r\u0131lmak \u00fczere harici kullan\u0131c\u0131lar taraf\u0131ndan zaten kullan\u0131lm\u0131\u015ft\u0131r.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"There_is_also_the_training_strategy\"><\/span><strong>Ayr\u0131ca <\/strong><strong>ve<\/strong><strong> e\u011fitim stratejisi.<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>Janus'un e\u011fitim a\u015famas\u0131 b\u00f6l\u00fcmlemesine uygun olarak, Janus Pro toplam \u00fc\u00e7 e\u011fitim a\u015famas\u0131na sahiptir ve makale bunlar\u0131 do\u011frudan A\u015fama I, A\u015fama II ve A\u015fama III olarak ay\u0131rmaktad\u0131r.<\/p>\n\n\n\n<p>Her bir a\u015faman\u0131n temel e\u011fitim fikirleri ve e\u011fitim hedefleri korunurken, Janus-Pro \u00fc\u00e7 a\u015famadaki e\u011fitim s\u00fcresi ve e\u011fitim verilerinde iyile\u015ftirmeler yapm\u0131\u015ft\u0131r. A\u015fa\u011f\u0131da \u00fc\u00e7 a\u015famadaki spesifik iyile\u015ftirmeler yer almaktad\u0131r:<\/p>\n\n\n\n<h4 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"Stage_I_%E2%80%93_Longer_training_time\"><\/span><strong>A\u015fama I - Daha uzun e\u011fitim s\u00fcresi<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h4>\n\n\n\n<p>Janus ile kar\u015f\u0131la\u015ft\u0131r\u0131ld\u0131\u011f\u0131nda Janus-Pro, \u00f6zellikle g\u00f6rsel k\u0131s\u0131mdaki adapt\u00f6rlerin ve g\u00f6r\u00fcnt\u00fc ba\u015fl\u0131klar\u0131n\u0131n e\u011fitiminde A\u015fama I'deki e\u011fitim s\u00fcresini uzatm\u0131\u015ft\u0131r. Bu, g\u00f6rsel \u00f6zelliklerin \u00f6\u011frenilmesine daha fazla e\u011fitim s\u00fcresi verildi\u011fi anlam\u0131na gelir ve modelin g\u00f6r\u00fcnt\u00fclerin ayr\u0131nt\u0131l\u0131 \u00f6zelliklerini (piksel-anlamsal e\u015fleme gibi) tam olarak anlayabilece\u011fi umulmaktad\u0131r.<\/p>\n\n\n\n<p>Bu geni\u015fletilmi\u015f e\u011fitim, g\u00f6rsel b\u00f6l\u00fcm\u00fcn e\u011fitiminin di\u011fer mod\u00fcller taraf\u0131ndan bozulmamas\u0131n\u0131 sa\u011flamaya yard\u0131mc\u0131 olur.<\/p>\n\n\n\n<h4 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"Stage_II_%E2%80%93_Removing_ImageNet_data_and_adding_multi-modal_data\"><\/span><strong>A\u015fama II - ImageNet verilerinin kald\u0131r\u0131lmas\u0131 ve \u00e7ok modlu verilerin eklenmesi<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h4>\n\n\n\n<p>A\u015fama II'de Janus daha \u00f6nce PixArt'a ba\u015fvurmu\u015f ve iki b\u00f6l\u00fcm halinde e\u011fitilmi\u015ftir. \u0130lk b\u00f6l\u00fcm g\u00f6r\u00fcnt\u00fc s\u0131n\u0131fland\u0131rma g\u00f6revi i\u00e7in ImageNet veri k\u00fcmesi kullan\u0131larak e\u011fitildi ve ikinci b\u00f6l\u00fcm normal metinden g\u00f6r\u00fcnt\u00fcye veriler kullan\u0131larak e\u011fitildi. A\u015fama II'deki zaman\u0131n yakla\u015f\u0131k \u00fc\u00e7te ikisi ilk k\u0131s\u0131mda e\u011fitime harcanm\u0131\u015ft\u0131r.<\/p>\n\n\n\n<p>Janus-Pro, A\u015fama II'deki ImageNet e\u011fitimini kald\u0131r\u0131r. Bu tasar\u0131m, modelin A\u015fama II e\u011fitimi s\u0131ras\u0131nda metinden g\u00f6r\u00fcnt\u00fcye verilere odaklanmas\u0131n\u0131 sa\u011flar. Deneysel sonu\u00e7lara g\u00f6re bu, metinden g\u00f6r\u00fcnt\u00fcye verilerin kullan\u0131m\u0131n\u0131 \u00f6nemli \u00f6l\u00e7\u00fcde art\u0131rabilir.<\/p>\n\n\n\n<p>E\u011fitim y\u00f6ntemi tasar\u0131m\u0131n\u0131n ayarlanmas\u0131na ek olarak, A\u015fama II'de kullan\u0131lan e\u011fitim veri seti art\u0131k tek bir g\u00f6r\u00fcnt\u00fc s\u0131n\u0131fland\u0131rma g\u00f6reviyle s\u0131n\u0131rl\u0131 de\u011fildir, ayn\u0131 zamanda ortak e\u011fitim i\u00e7in g\u00f6r\u00fcnt\u00fc a\u00e7\u0131klamas\u0131 ve diyalog gibi di\u011fer \u00e7ok modlu veri t\u00fcrlerini de i\u00e7erir.<\/p>\n\n\n\n<h4 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"Stage_III_%E2%80%93_Optimizing_the_data_ratio\"><\/span><strong>A\u015fama III - Veri oran\u0131n\u0131n optimize edilmesi<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h4>\n\n\n\n<p>A\u015fama III e\u011fitiminde, Janus-Pro farkl\u0131 e\u011fitim verisi t\u00fcrlerinin oran\u0131n\u0131 ayarlar.<\/p>\n\n\n\n<p>Daha \u00f6nce, Janus taraf\u0131ndan A\u015fama III'te kullan\u0131lan e\u011fitim verilerindeki \u00e7ok modlu anlama verileri, d\u00fcz metin verileri ve metin-imaj verilerinin oran\u0131 7:3:10 idi. Janus-Pro, son iki veri t\u00fcr\u00fcn\u00fcn oran\u0131n\u0131 azalt\u0131r ve \u00fc\u00e7 veri t\u00fcr\u00fcn\u00fcn oran\u0131n\u0131 5:1:4 olarak ayarlar, yani \u00e7ok modlu anlama g\u00f6revine daha fazla \u00f6nem verir.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"Lets_look_at_the_training_data\"><\/span><strong>E\u011fitim verilerine bakal\u0131m.<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n\n<p>Janus ile kar\u015f\u0131la\u015ft\u0131r\u0131ld\u0131\u011f\u0131nda, Janus-Pro bu kez y\u00fcksek kaliteli<strong> sentetik veri.<\/strong><\/p>\n\n\n\n<p>\u00c7ok modlu anlama ve g\u00f6r\u00fcnt\u00fc olu\u015fturma i\u00e7in e\u011fitim verilerinin miktar\u0131n\u0131 ve \u00e7e\u015fitlili\u011fini art\u0131r\u0131r.<\/p>\n\n\n\n<h4 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"Expansion_of_multimodal_understanding_data\"><\/span><strong>\u00c7ok modlu anlama verilerinin geni\u015fletilmesi:<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h4>\n\n\n\n<p>Janus-Pro, e\u011fitim s\u0131ras\u0131nda DeepSeek-VL2 veri k\u00fcmesine ba\u015fvurur ve yaln\u0131zca g\u00f6r\u00fcnt\u00fc a\u00e7\u0131klama veri k\u00fcmelerini de\u011fil, ayn\u0131 zamanda tablolar, grafikler ve belgeler gibi karma\u015f\u0131k sahnelerin veri k\u00fcmelerini de i\u00e7eren yakla\u015f\u0131k 90 milyon ek veri noktas\u0131 ekler.<\/p>\n\n\n\n<p>Denetimli ince ayar a\u015famas\u0131 s\u0131ras\u0131nda (A\u015fama III), MEME anlay\u0131\u015f\u0131 ve diyalog (\u00c7ince diyalog dahil) deneyimi iyile\u015ftirme ile ilgili veri k\u00fcmeleri eklemeye devam eder.<\/p>\n\n\n\n<h4 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"Expansion_of_visual_generation_data\"><\/span><strong>G\u00f6rsel \u00fcretim verilerinin geni\u015fletilmesi:<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h4>\n\n\n\n<p>Orijinal ger\u00e7ek d\u00fcnya verilerinin kalitesi d\u00fc\u015f\u00fck ve g\u00fcr\u00fclt\u00fc seviyeleri y\u00fcksekti, bu da modelin karars\u0131z \u00e7\u0131kt\u0131lar ve metinden g\u00f6r\u00fcnt\u00fcye g\u00f6revlerinde yetersiz estetik kalitede g\u00f6r\u00fcnt\u00fcler \u00fcretmesine neden oldu.<\/p>\n\n\n\n<p>Janus-Pro, e\u011fitim a\u015famas\u0131na yakla\u015f\u0131k 72 milyon yeni y\u00fcksek estetikli sentetik veri ekleyerek \u00f6n e\u011fitim a\u015famas\u0131ndaki ger\u00e7ek verilerin sentetik verilere oran\u0131n\u0131 1:1'e getirmi\u015ftir.<\/p>\n\n\n\n<p>Sentetik verilere ili\u015fkin istemlerin t\u00fcm\u00fc kamu kaynaklar\u0131ndan al\u0131nm\u0131\u015ft\u0131r. Deneyler, bu verilerin eklenmesinin modelin daha h\u0131zl\u0131 yak\u0131nsamas\u0131n\u0131 sa\u011flad\u0131\u011f\u0131n\u0131 ve \u00fcretilen g\u00f6r\u00fcnt\u00fclerin kararl\u0131l\u0131k ve g\u00f6rsel g\u00fczellik a\u00e7\u0131s\u0131ndan belirgin iyile\u015ftirmelere sahip oldu\u011funu g\u00f6stermi\u015ftir.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"The_continuation_of_an_efficiency_revolution\"><\/span>Bir verimlilik devriminin devam\u0131 m\u0131?<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p>Genel olarak, bu s\u00fcr\u00fcmle DeepSeek g\u00f6rsel modellere verimlilik devrimini getirmi\u015ftir.<\/p>\n\n\n\n<p>Tek bir i\u015fleve odaklanan g\u00f6rsel modellerin veya belirli bir g\u00f6revi tercih eden \u00e7ok modlu modellerin aksine, Janus-Pro g\u00f6r\u00fcnt\u00fc olu\u015fturma ve \u00e7ok modlu anlama gibi iki ana g\u00f6revin etkilerini ayn\u0131 modelde dengeler.<\/p>\n\n\n\n<p>Ayr\u0131ca, k\u00fc\u00e7\u00fck parametrelerine ra\u011fmen, de\u011ferlendirmedeOpenAI DALL-E 3 ve SD3-Medium'u geride b\u0131rakt\u0131.<\/p>\n\n\n\n<p>Yere kadar geni\u015fletildi\u011finde, i\u015fletmenin yaln\u0131zca g\u00f6r\u00fcnt\u00fc olu\u015fturma ve anlaman\u0131n iki i\u015flevini do\u011frudan uygulamak i\u00e7in bir model da\u011f\u0131tmas\u0131 gerekir. Sadece 7B'l\u0131k bir boyutla birle\u015fti\u011finde, da\u011f\u0131t\u0131m zorlu\u011fu ve maliyeti \u00e7ok daha d\u00fc\u015f\u00fckt\u00fcr.<\/p>\n\n\n\n<p>R1 ve V3'\u00fcn \u00f6nceki s\u00fcr\u00fcmleriyle ba\u011flant\u0131l\u0131 olarak DeepSeek, oyunun mevcut kurallar\u0131na \u015fu \u015fekilde meydan okuyor<strong> \"kompakt mimari yenilik, hafif modeller, a\u00e7\u0131k kaynak modeller ve ultra d\u00fc\u015f\u00fck e\u011fitim maliyetleri\"<\/strong>. Bat\u0131l\u0131 teknoloji devleri ve hatta Wall Street aras\u0131ndaki pani\u011fin nedeni budur.<\/p>\n\n\n\n<p>Birka\u00e7 g\u00fcnd\u00fcr kamuoyunu pe\u015finden s\u00fcr\u00fckleyen Sam Altman, nihayet X'teki DeepSeek ile ilgili bilgilere olumlu yan\u0131t verdi - R1'i \u00f6verken, OpenAI'nin baz\u0131 duyurular yapaca\u011f\u0131n\u0131 s\u00f6yledi.<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1051\" height=\"1280\" src=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/ade0e532-b451-4eff-832f-dcf20cff8f64.png\" alt=\"\" class=\"wp-image-868\" srcset=\"https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/ade0e532-b451-4eff-832f-dcf20cff8f64.png 1051w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/ade0e532-b451-4eff-832f-dcf20cff8f64-246x300.png 246w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/ade0e532-b451-4eff-832f-dcf20cff8f64-841x1024.png 841w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/ade0e532-b451-4eff-832f-dcf20cff8f64-768x935.png 768w, https:\/\/janusai.pro\/wp-content\/uploads\/2025\/02\/ade0e532-b451-4eff-832f-dcf20cff8f64-10x12.png 10w\" sizes=\"auto, (max-width: 1051px) 100vw, 1051px\" \/><\/figure>","protected":false},"excerpt":{"rendered":"<p>DeepSeek web sitesini g\u00fcncelledi. Y\u0131lba\u015f\u0131 gecesinin erken saatlerinde DeepSeek aniden GitHub'da Janus proje alan\u0131n\u0131n Janus-Pro modeli ve teknik raporunun kayna\u011f\u0131n\u0131 a\u00e7t\u0131\u011f\u0131n\u0131 duyurdu. \u0130lk olarak, birka\u00e7 \u00f6nemli noktay\u0131 vurgulayal\u0131m: Sekt\u00f6r\u00fcn zor sorununu \u00e7\u00f6zmeyi ama\u00e7l\u0131yor: \u00e7ok modlu anlama ve g\u00f6r\u00fcnt\u00fc olu\u015fturmay\u0131 dengelemek...<\/p>","protected":false},"author":2,"featured_media":704,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_kadence_starter_templates_imported_post":false,"_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,"footnotes":""},"categories":[1],"tags":[],"class_list":["post-857","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/janusai.pro\/tr\/wp-json\/wp\/v2\/posts\/857","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/janusai.pro\/tr\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/janusai.pro\/tr\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/janusai.pro\/tr\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/janusai.pro\/tr\/wp-json\/wp\/v2\/comments?post=857"}],"version-history":[{"count":1,"href":"https:\/\/janusai.pro\/tr\/wp-json\/wp\/v2\/posts\/857\/revisions"}],"predecessor-version":[{"id":870,"href":"https:\/\/janusai.pro\/tr\/wp-json\/wp\/v2\/posts\/857\/revisions\/870"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/janusai.pro\/tr\/wp-json\/wp\/v2\/media\/704"}],"wp:attachment":[{"href":"https:\/\/janusai.pro\/tr\/wp-json\/wp\/v2\/media?parent=857"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/janusai.pro\/tr\/wp-json\/wp\/v2\/categories?post=857"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/janusai.pro\/tr\/wp-json\/wp\/v2\/tags?post=857"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}