{"id":90100,"date":"2025-10-21T15:06:32","date_gmt":"2025-10-21T12:06:32","guid":{"rendered":"https:\/\/forklog.com\/en\/?p=90100"},"modified":"2025-10-21T15:10:35","modified_gmt":"2025-10-21T12:10:35","slug":"deepseek-unveils-text-compression-technology-for-ai","status":"publish","type":"post","link":"https:\/\/forklog.com\/en\/deepseek-unveils-text-compression-technology-for-ai\/","title":{"rendered":"DeepSeek Unveils Text Compression Technology for AI"},"content":{"rendered":"<p>Chinese AI startup DeepSeek has <a href=\"https:\/\/deepseekocr.app\/\">introduced<\/a> a new multimodal AI capable of processing large and complex documents using significantly fewer tokens.<\/p>\n<p>DeepSeek-OCR employs visual perception as a means of compressing information.<\/p>\n<p>The system is the result of research into the &#8220;role of <span data-descr=\"converts an image into a numerical representation understandable by the model\" class=\"old_tooltip\">visual encoders<\/span>&#8221; for text compression in large language models (LLM). This approach enables neural networks to handle vast amounts of information without a proportional increase in computational costs.<\/p>\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>\u201cWith DeepSeek-OCR, we demonstrated that compressing text through visual representations allows for a 7\u201320 fold reduction in tokens at various stages of context. This opens a promising direction for addressing the long history problem in LLMs,\u201d the company stated.<\/p>\n<\/blockquote>\n<p>DeepSeek-OCR consists of two main components:<\/p>\n<ul class=\"wp-block-list\">\n<li>DeepEncoder \u2014 the encoder;<\/li>\n<li>DeepSeek3B-MoE-A570M \u2014 the decoder.<\/li>\n<\/ul>\n<p>The first serves as the main computational core of the model. It maintains low activity while processing high-resolution images, achieving a substantial level of compression. This reduces the number of tokens.<\/p>\n<p>The decoder, a <span data-descr=\"a neural network architecture where the model consists not of a single large block, but of a set of 'experts' \u2014 individual subnetworks, each specializing in its own type of tasks or data\" class=\"old_tooltip\">Mixture-of-Experts<\/span> model with 570 million parameters, is responsible for restoring the original text. The architecture divides the neural network into several independent subnetworks \u2014 &#8220;experts,&#8221; each specializing in its part of the input data. Together, they solve the overall task.<\/p>\n<p>DeepSeek-OCR can analyze complex structured visual content, tables, formulas, and geometric diagrams. According to the company, this makes the model particularly useful for applications in finance and scientific research.<\/p>\n<p>The company noted that DeepSeek-OCR achieved 97% decoding accuracy. At a 20x compression ratio, the model retained about 60%. This underscores its ability to preserve information even at extreme levels of compression.<\/p>\n<p>On OmniDocBench \u2014 a benchmark test for evaluating the understanding of diverse documents \u2014 DeepSeek-OCR outperformed leading optical character recognition models like GOT-OCR 2.0 and MinerU 2.0, while using significantly fewer tokens.<\/p>\n<p>Back in August, the startup <a href=\"https:\/\/forklog.com\/en\/news\/deepseek-unveils-updated-ai-model-v3-1\">updated<\/a> its flagship AI model V3.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>DeepSeek has introduced a new multimodal AI capable of processing large and complex documents using significantly fewer tokens.<\/p>\n","protected":false},"author":1,"featured_media":90101,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"select":"1","news_style_id":"1","cryptorium_level":"","_short_excerpt_text":"DeepSeek unveils AI for processing documents with fewer tokens.","creation_source":"","_metatest_mainpost_news_update":false,"footnotes":""},"categories":[3],"tags":[438,1743],"class_list":["post-90100","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news-and-analysis","tag-artificial-intelligence","tag-deepseek"],"aioseo_notices":[],"amp_enabled":true,"views":"216","promo_type":"1","layout_type":"1","short_excerpt":"DeepSeek unveils AI for processing documents with fewer tokens.","is_update":"","_links":{"self":[{"href":"https:\/\/forklog.com\/en\/wp-json\/wp\/v2\/posts\/90100","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/forklog.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/forklog.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/forklog.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/forklog.com\/en\/wp-json\/wp\/v2\/comments?post=90100"}],"version-history":[{"count":1,"href":"https:\/\/forklog.com\/en\/wp-json\/wp\/v2\/posts\/90100\/revisions"}],"predecessor-version":[{"id":90102,"href":"https:\/\/forklog.com\/en\/wp-json\/wp\/v2\/posts\/90100\/revisions\/90102"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/forklog.com\/en\/wp-json\/wp\/v2\/media\/90101"}],"wp:attachment":[{"href":"https:\/\/forklog.com\/en\/wp-json\/wp\/v2\/media?parent=90100"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/forklog.com\/en\/wp-json\/wp\/v2\/categories?post=90100"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/forklog.com\/en\/wp-json\/wp\/v2\/tags?post=90100"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}