DeepSeek V3.2 正式版发布，V4 还没来，但已经是开源模型里 Agent 能力最强了

Core Insights - DeepSeek has released the official version of its V3.2 model, which significantly enhances reasoning and agent capabilities compared to previous versions [2][9] - The V3.2-Speciale version is an open-source model that performs comparably to Gemini-3.0-Pro on mainstream reasoning benchmarks and has achieved gold medal levels in several prestigious competitions [3][11] - The integration of the DeepSeek Sparse Attention (DSA) technology in V3.2 improves long text processing efficiency and reduces costs by over 50% [3][10] Model Development - The V3 series has been iterated over the past year, with V3.2 being the latest release, focusing on unifying thinking and non-thinking models, a trend seen in other closed-source models like Gemini and GPT-5 [6][9] - The release timeline for DeepSeek models in 2025 includes various versions, each with specific enhancements, such as the introduction of DSA in V3.2 for stability and reasoning improvements [7][8] Performance Metrics - DeepSeek-V3.2 has achieved reasoning capabilities on par with GPT-5 and has shown significant improvements in output length and computational efficiency compared to Kimi-K2-Thinking [10][14] - The V3.2-Speciale version excels in complex tasks, achieving high scores in various academic competitions, including IMO 2025 and ICPC 2025, with notable rankings among human competitors [11][14] Tool Utilization - A key advancement in V3.2 is the incorporation of thinking processes into tool calls, allowing the model to support both thinking and non-thinking modes in its operations [15][18] - DeepSeek has developed a large-scale agent training data synthesis method that enhances the model's generalization capabilities by creating numerous "hard-to-answer, easy-to-verify" tasks [16][18]