[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-638899-105":3,"detail-sidebar-cat-0-en-105":81,"doc-detail-638899-en":130},{"code":4,"msg":5,"data":6},0,"ok",{"site_id":7,"language":8,"slug":9,"title":10,"keywords":11,"description":12,"schema_data":13,"social_meta":74,"head_meta":76,"extra_data":78,"updated_unix":80},105,"en","evaluation-of-deepseek-ai-models-executive-summary","Evaluation of DeepSeek AI Models - Executive Summary","","Executive summary of a CAISI assessment at NIST evaluating three DeepSeek models (R1, R1-0528, V3.1) against four U.S. reference models (including GPT-5 variants and Anthropic’s Opus 4) across 19 benchmarks. Findings indicate DeepSeek trails leading U.S. systems on performance, with the largest gaps in software engineering and cyber tasks, and higher usage costs. Security results show substantially greater susceptibility to agent hijacking and jailbreaking. The report also finds more frequent amplification of CCP narratives and notes rapid PRC model adoption following the DeepSeek R1 release, raising risks for developers, consumers, and U.S. national security.",{"@graph":14,"@context":73},[15,34,56],{"@type":16,"itemListElement":17},"BreadcrumbList",[18,23,27,31],{"item":19,"name":20,"@type":21,"position":22},"https://docshare.wps.com","Home","ListItem",1,{"item":24,"name":25,"@type":21,"position":26},"https://docshare.wps.com/document/","Document",2,{"item":28,"name":29,"@type":21,"position":30},"https://docshare.wps.com/document/research-report/","Research & Report",3,{"item":32,"name":10,"@type":21,"position":33},"https://docshare.wps.com/document/evaluation-of-deepseek-ai-models-executive-summary/638899/",4,{"url":32,"name":10,"@type":35,"image":36,"author":41,"headline":10,"publisher":44,"fileFormat":47,"inLanguage":8,"description":12,"dateModified":48,"datePublished":49,"encodingFormat":47,"isAccessibleForFree":50,"interactionStatistic":51},"DigitalDocument",{"url":37,"@type":38,"width":39,"height":40},"https://docshare.wps.com/thumbnails/evaluation-of-deepseek-ai-models-executive-summary/638899.png","ImageObject",300,407,{"name":42,"@type":43},"Valentina","Person",{"url":19,"name":45,"@type":46},"DocShare","Organization","application/pdf","2026-11-03","2026-10-22",true,{"@type":52,"interactionType":53,"userInteractionCount":55},"InteractionCounter",{"@type":54},"ViewAction",7,{"@type":57,"mainEntity":58},"FAQPage",[59,65,69],{"name":60,"@type":61,"acceptedAnswer":62},"How do DeepSeek models compare with U.S. reference models on performance?","Question",{"text":63,"@type":64},"DeepSeek performance lags behind the best U.S. reference models across almost every benchmark. The gap is largest for software engineering and cyber tasks, where the best U.S. model solves 20–80% more tasks.","Answer",{"name":66,"@type":61,"acceptedAnswer":67},"What did the evaluation find about cost efficiency?",{"text":68,"@type":64},"DeepSeek models cost more to use than comparable U.S. models. One U.S. reference model was 35% less expensive on average to reach a similar level across 13 tested performance benchmarks.",{"name":70,"@type":61,"acceptedAnswer":71},"What security weaknesses were observed in DeepSeek models?",{"text":72,"@type":64},"DeepSeek models were more susceptible to agent hijacking and jailbreaking. Agents built on R1-0528 were about 12 times more likely than U.S. frontier models to follow malicious instructions, and R1-0528 complied with 94% of overtly malicious jailbreaking requests versus 8% for U.S. references.","https://schema.org",{"og:url":32,"og:type":75,"og:title":10,"og:site_name":45,"og:description":12},"article",{"robots":77,"canonical":32},"index,follow",{"doc_id":79,"site_id":7},638899,1792856582,{"code":4,"msg":82,"data":83},"success",[84,88,92,96,101,106,110,114,119,122,126],{"id":22,"doc_module":4,"doc_module_name":25,"category_name":85,"show_sort_weight":86,"slug":87},"Story & Novel",90,"story-novel",{"id":26,"doc_module":4,"doc_module_name":25,"category_name":89,"show_sort_weight":90,"slug":91},"Literature",80,"literature",{"id":33,"doc_module":4,"doc_module_name":25,"category_name":93,"show_sort_weight":94,"slug":95},"Exam",70,"exam",{"id":97,"doc_module":4,"doc_module_name":25,"category_name":98,"show_sort_weight":99,"slug":100},5,"Comic",60,"comic",{"id":102,"doc_module":4,"doc_module_name":25,"category_name":103,"show_sort_weight":104,"slug":105},6,"Technology",50,"technology",{"id":55,"doc_module":4,"doc_module_name":25,"category_name":107,"show_sort_weight":108,"slug":109},"Healthcare",40,"healthcare",{"id":111,"doc_module":4,"doc_module_name":25,"category_name":29,"show_sort_weight":112,"slug":113},8,30,"research-report",{"id":115,"doc_module":4,"doc_module_name":25,"category_name":116,"show_sort_weight":117,"slug":118},9,"Religion & Spirituality",20,"religion-spirituality",{"id":117,"doc_module":4,"doc_module_name":25,"category_name":120,"show_sort_weight":117,"slug":121},"World Cup","world-cup",{"id":123,"doc_module":4,"doc_module_name":25,"category_name":124,"show_sort_weight":123,"slug":125},10,"Lifestyle","lifestyle",{"id":127,"doc_module":4,"doc_module_name":25,"category_name":128,"show_sort_weight":97,"slug":129},19,"General","general",{"code":4,"msg":82,"data":131},{"doc_id":79,"user_id":132,"nickname":42,"user_avatar":133,"doc_module":4,"category_id":111,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":55,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":139,"language":140,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":141,"faqs":142,"seo_title":143,"seo_description":12,"update_tm":144,"read_time":145},13056703020460,"https://ap-avatar.wpscdn.com/avatar/be000253dac470eee5d?_k=1778207105932848923","Evaluation of DeepSeek AI Models  \nCenter for AI Standards and Innovation National Institute of Standards and Technology  \n1. Executive Summary  \nPresident Trump, through his AI Action Plan, and Secretary of Commerce Howard Lutnick have tasked the Center for AI Standards and Innovation (CAISI) at the National Institute of Standards and Technology (NIST) with assessing the capabilities of U.S. and adversary AI systems, the adoption of foreign AI systems, and the state of international AI competition.  \nActing on these directives, in September 2025 CAISI conducted a technical evaluation of three DeepSeek models (R1, R1-0528, and V3 . 1) and four U.S. reference models (OpenAI’s GPT-5, GPT-5-mini, gpt-oss, and Anthropic’s Opus 4) on 19 benchmarks across a range of domains. These evaluations include state-of-the-art public benchmarks as well as private benchmarks built by CAISI staff in partnership with academic institutions and other federal agencies.  \nCAISI’s evaluation of DeepSeek models yielded six key findings:  \nDeepSeek performance lags behind the best U.S. reference models.  \nThe best U.S. model outperforms the best DeepSeek model (DeepSeek V3 . 1) across almost every benchmark. The gap is largest for software engineering and cyber tasks, where the best U.S. model solves 20-80% more tasks than the best DeepSeek model. V3 .1, DeepSeek’s most recent model, nevertheless outperforms DeepSeek’s earlier R1 models.  \nDeepSeek models cost more to use than comparable U.S. models.  \nOne U.S. reference model cost 35% less on average than the best DeepSeek model to perform at a similar level across all 13 performance benchmarks tested.  \nDeepSeek models are far more susceptible to agent hijacking attacks than frontier U.S. models.  \nAgents based on DeepSeek’s most secure model (R1-0528) were, on average, 12 times likelier than evaluated U.S. frontier models (GPT-5 and Opus 4) to follow malicious instructions designed to derail them from user tasks.  \nDeepSeek models are far more susceptible to jailbreaking attacks than U.S. models.  \nDeepSeek’s most secure model (R1-0528) complied with 94% of overtly malicious requests that used common jailbreaking techniques, compared to 8% of requests for U.S. reference models.  \nDeepSeek models advance Chinese Communist Party (CCP) narratives much more frequently than U.S. models.  \nOn a dataset of politically sensitive questions for the CCP, on average, DeepSeek models echoed 4 times as many inaccurate and misleading CCP narratives as U.S. reference models did.  \nAdoption of PRC models has greatly increased since DeepSeek R1 was released.  \nThe release of DeepSeek R1 has driven adoption of PRC models across the AI ecosystem. Downloads of DeepSeek models on model sharing platforms have increased nearly 1000% since January.  \n***  \nOverall, while DeepSeek’s models lag behind leading U.S. models, DeepSeek remains a leading open-weight model developer and has contributed to a rapid increase in adoption of PRC models globally. Given the security shortcomings and censorship of the PRC models evaluated in this report, the expanding use of these models may pose a risk to application developers, to consumers, and to U.S. national security.  \nTable of Contents  \n1. Executive Summary............................................................................................................................................ 2  \n2. Methodology ..................................................................................................................................................... 4  \n3. Overview of Findings ......................................................................................................................................... 6  \n3.1 Performance Evaluations Overview ..................................................................................................................... 7  \n3.2 Cost Efficiency Analysis Overview ..............................................................","cbCaisJv50DEerNc","https://ap.wps.com/l/cbCaisJv50DEerNc","pdf",4784992,69,"English","# 1. Executive Summary\n# 2. Methodology\n# 3. Overview of Findings\n## 3.1 Performance Evaluations Overview\n## 3.2 Cost Efficiency Analysis Overview\n## 3.3 Security Evaluations Overview\n## 3.4 Chinese Communist Party Censorship Evaluation Overview\n## 3.5 Model Adoption Overview\n# 4. Performance Evaluations\n## 4.1 Cyber\n## 4.2 Software Development\n## 4.3 Science and Knowledge\n## 4.4 Mathematical Reasoning\n# 5. Cost Efficiency Analysis\n# 6. Security Evaluations\n## 6.1 Agent Hijacking\n## 6.2 Jailbreaking","[{\"question\":\"How do DeepSeek models compare with U.S. reference models on performance?\",\"answer\":\"DeepSeek performance lags behind the best U.S. reference models across almost every benchmark. The gap is largest for software engineering and cyber tasks, where the best U.S. model solves 20–80% more tasks.\"},{\"question\":\"What did the evaluation find about cost efficiency?\",\"answer\":\"DeepSeek models cost more to use than comparable U.S. models. One U.S. reference model was 35% less expensive on average to reach a similar level across 13 tested performance benchmarks.\"},{\"question\":\"What security weaknesses were observed in DeepSeek models?\",\"answer\":\"DeepSeek models were more susceptible to agent hijacking and jailbreaking. Agents built on R1-0528 were about 12 times more likely than U.S. frontier models to follow malicious instructions, and R1-0528 complied with 94% of overtly malicious jailbreaking requests versus 8% for U.S. references.\"}]","Evaluation of DeepSeek AI Models - Executive Summary | PDF",1792700302,174]