Chinese artificial intelligence company DeepSeek has introduced two new models focused on reasoning, demonstrating that open-source systems from China are now matching and, in some cases, surpassing the performance of leading US AI developers in mathematics and coding assessments.<\/p>\n
The Hangzhou-based company announced DeepSeek-V3.2 and the more advanced DeepSeek-V3.2-Speciale on December 1.<\/p>\n
According to DeepSeek, both open-source versions offer reasoning results on par with OpenAI’s GPT-5 and Google’s Gemini 3 Pro.<\/p>\n
The Speciale version achieved gold-standard performance in four major competitions: the International Mathematical Olympiad, the Chinese Mathematical Olympiad, the ICPC World Finals, and the 2025 International Olympiad in Informatics.<\/p>\n
Surpassed GPT-5 and Gemini in mathematics test<\/strong><\/p>\n In the 2025 American Invitational Mathematics Examination (AIME), DeepSeek-V3.2-Speciale scored 96.0%.<\/p>\n This result is higher than GPT-5 High’s score of 94.6% and Gemini 3 Pro’s score of 95.0%.<\/p>\n The model achieved a score of 73.1% on the SWE Verified test, a key benchmark for coding proficiency, but this score was behind Gemini 3 Pro’s 76.2%.<\/p>\n China’s open-source push is strengthening<\/strong><\/p>\n This development marks another step in China’s effort to challenge US leadership in artificial intelligence through open-source innovation.<\/p>\n Founded in 2023 by Liang Wenfeng and backed by the quantitative hedge fund High-Flyer, DeepSeek is positioning itself as a producer of high-performance, low-cost AI systems that can compete with Silicon Valley’s proprietary systems.<\/p>\n Bloomberg noted that DeepSeek-V3.2 is the successor to the previous V3.2-Exp model introduced in September. The new model is now accessible through DeepSeek’s application, website, and API.<\/p>\n The standard V3.2 is designed as a “daily use model” offering GPT-5 level capabilities with strong efficiency, while the V3.2-Speciale version is optimized for maximum reasoning power on highly complex problems.<\/p>\n China takes the lead in open-source downloads<\/strong><\/p>\n Recent findings show that China has surpassed the US in open-source artificial intelligence model downloads.<\/p>\n Research conducted by MIT and Hugging Face revealed that Chinese models now account for 17% of global downloads, while the share of American models is 15.8%.<\/p>\n Analysts attribute this shift to China’s rapid release cycles and its emphasis on lightweight, accessible models that perform well on less advanced hardware.<\/p>\n This approach has been partly shaped by US export restrictions on high-quality AI chips.<\/p>\n Competition is accelerating within China<\/strong><\/p>\n Meanwhile, China’s AI industry is becoming increasingly competitive. Other major players like Alibaba and Baidu are also rapidly developing their own systems.<\/p>\n Alibaba’s Qwen 3 model recently achieved excellent results on the AIME 2025 and several other mathematics tests.<\/p>\n According to Stanford’s\u00a02025 Artificial Intelligence Index Report<\/em>, China is responsible for approximately 70% of global AI-related patent applications.<\/p>\n ‘Reasoning-first models built for agents’<\/strong><\/p>\n DeepSeek’s technical documentation describes the V3.2 family as “reasoning-first models built for agents.”<\/p>\n The models are supported by a new training methodology that synthesizes over 85,000 difficult tasks across 1,800 different environments.<\/p>\n The company also states that V3.2 is the first model to embed reasoning directly into tool-use workflows.<\/p>\n This feature allows the model to switch between reasoning and non-reasoning modes when using external tools such as code executors or search assistants.<\/p>\n The V3.2-Speciale model, which will be accessible only via API until December 15, requires more processing resources but is designed to deliver what DeepSeek calls “fully maximized reasoning capacity.”<\/p>\n Both variants include the DeepSeek Sparse Attention method, a technique aimed at reducing the computational cost of processing long inputs by roughly 50-70%.<\/p>\n","protected":false},"excerpt":{"rendered":" Chinese artificial intelligence company DeepSeek has introduced two new models focused on reasoning, demonstrating that open-source systems from China are now matching and, in some cases, surpassing the performance of leading US AI developers in mathematics and coding assessments. The Hangzhou-based company announced DeepSeek-V3.2 and the more advanced DeepSeek-V3.2-Speciale on December 1. According to DeepSeek, […]<\/p>\n","protected":false},"author":1,"featured_media":12429,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ep_exclude_from_search":false,"footnotes":""},"categories":[4],"tags":[5443,14585,5661,14586,2854,2122,2856,529,14584,5442,14579,14580,73,11939,503,14583,14581,4371,14582,821],"class_list":["post-12428","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-asia","tag-ai-chips","tag-ai-leadership","tag-ai-model","tag-aime","tag-alibaba","tag-artificial-intelligence","tag-baidu","tag-china","tag-coding-proficiency","tag-deepseek","tag-deepseek-v3-2","tag-gemini-3-pro","tag-google","tag-gpt-5","tag-headline","tag-mathematics-benchmarks","tag-open-source-ai","tag-openai","tag-reasoning","tag-us"],"yoast_head":"\n