DeepSeek 重大发布

Core Insights - DeepSeek has released two official model versions: DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, with the former available on the official website, app, and API, while the latter is currently accessible only as a temporary API for community evaluation [1][3]. Model Performance - DeepSeek-V3.2 aims to balance reasoning capability and output length, making it suitable for daily use. In benchmark tests, it achieved performance comparable to GPT-5 and slightly below Gemini-3.0-Pro, with a significant reduction in output length compared to Kimi-K2-Thinking, leading to lower computational costs and reduced user wait times [3][4]. - DeepSeek-V3.2-Speciale is designed to push the limits of reasoning capabilities, serving as an enhanced version of DeepSeek-V3.2, and incorporates theorem-proving abilities from DeepSeek-Math-V2. It performed comparably to Gemini-3.0-Pro in mainstream reasoning benchmarks and won gold medals in several prestigious competitions, including IMO 2025 and ICPC World Finals 2025, achieving second and tenth place among human competitors, respectively [3][4]. Benchmark Comparisons - In various benchmark tests, DeepSeek-V3.2 and DeepSeek-V3.2-Speciale demonstrated competitive performance: - AIME 2025: DeepSeek-V3.2 scored 93.1, while DeepSeek-V3.2-Speciale scored 96.0 [4]. - HMMT Feb 2025: DeepSeek-V3.2 scored 92.5, and DeepSeek-V3.2-Speciale scored 99.2 [4]. - IMOAnswerBench: DeepSeek-V3.2 scored 78.3, and DeepSeek-V3.2-Speciale scored 84.5 [4]. - CodeForces: DeepSeek-V3.2 scored 2386, while DeepSeek-V3.2-Speciale scored 2701 [4]. Cost Efficiency - The introduction of DeepSeek-V3.2-Exp, based on V3.1-Terminus with a new attention mechanism (DSA), has led to significant improvements in training and reasoning efficiency, resulting in a notable reduction in model costs. This cost reduction enhances the model's cost-effectiveness and potential for broader application [4].