国庆前搞大事！DeepSeek 新模型速度翻 3 倍，API 直接半价！网友调侃：这假没法休了

Core Viewpoint - DeepSeek has released its experimental version DeepSeek-V3.2-Exp, which significantly improves long text training and inference efficiency while maintaining output quality compared to its predecessor V3.1-Terminus [5][6]. Group 1: Model Performance - DeepSeek-V3.2-Exp introduces DeepSeek Sparse Attention (DSA), achieving a 2-3 times increase in long text inference speed and a 30%-40% reduction in memory usage, along with a 50% improvement in training efficiency [5]. - In benchmark tests, DeepSeek-V3.2-Exp performs comparably to V3.1-Terminus, with scores of 85.0 in MMLU-Pro and a slight improvement in AIME 2025, scoring 89.3 compared to 88.4 [5][6]. Group 2: Pricing Adjustments - Due to the reduced service costs associated with the new model, DeepSeek has lowered its API pricing by over 50%, with input prices dropping from 0.5 yuan to 0.2 yuan per million tokens for cache hits, and from 4 yuan to 2 yuan for cache misses. Output prices have decreased from 12 yuan to 3 yuan per million tokens [7].