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 being available on the official website, app, and API, while the latter is currently accessible only through a temporary API for community evaluation [2][4]. Model Performance - DeepSeek-V3.2 aims to balance reasoning capability and output length, achieving performance comparable to GPT-5 and slightly below Gemini-3.0-Pro in benchmark tests. It significantly reduces output length compared to Kimi-K2-Thinking, leading to lower computational costs and reduced user wait times [4]. - The Speciale version is designed to push the limits of reasoning capabilities, combining features from DeepSeek-V3.2 and DeepSeek-Math-V2, and has shown performance on par with Gemini-3.0-Pro in mainstream reasoning benchmarks [4]. Benchmark Results - In various benchmark tests, DeepSeek-V3.2-Speciale achieved notable results, including: - AIME 2025: 96.0 (23k) - HMMT Feb 2025: 99.2 (27k) - HMMT Nov 2025: 94.4 (25k) - IMOAnswerBench: 84.5 (45k) - CodeForces: 2701 (77k) - HILE: 30.6 (35k) [5]. Cost Efficiency - The introduction of the new attention mechanism DSA in DeepSeek-V3.2-Exp has led to significant improvements in training efficiency and a reduction in model costs, enhancing the model's cost-effectiveness and potential for broader application [5].