ChatGPT三周年遭DeepSeek暴击，23页技术报告藏着开源登顶的全部秘密

Core Insights - DeepSeek has launched two new models, DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, which significantly enhance reasoning capabilities, rivaling GPT-5 and Gemini-3.0-Pro, while addressing long-standing issues in open-source models [2][5][48]. Model Features - DeepSeek-V3.2 focuses on cost-effectiveness and daily use, achieving reasoning capabilities comparable to GPT-5, with faster and shorter outputs than Kimi-K2-Thinking, and introduces "thinking while using tools" [5][19]. - DeepSeek-V3.2-Speciale targets the upper limits of AI capabilities, performing exceptionally in competitions like IMO and ICPC, but is resource-intensive and does not support tool calls [5][19][38]. Technical Innovations - The introduction of DSA (Sparse Attention Mechanism) allows the model to focus on important parts of the input, significantly improving processing speed and efficiency, supporting a context length of 128K [9][12][13]. - DeepSeek invested over 10% of the pre-training budget in post-training resources, utilizing a stable and scalable reinforcement learning framework to enhance model performance [14][15]. Training Methodology - The training process involves "expert distillation" to create specialized models in various fields, followed by "mixed reinforcement learning training" to unify different task performances and prevent catastrophic forgetting [16][18]. - The model's performance is enhanced through a self-training pipeline, where AI generates and verifies its own training data across over 18,000 tasks, promoting self-evolution [30][32]. Performance Metrics - In benchmark tests, DeepSeek-V3.2 shows competitive performance with GPT-5 and Kimi-K2-Thinking across various metrics, while the Speciale version approaches or exceeds Gemini-3.0-Pro [33][34]. - The model achieved notable results in competitions, including gold medals in IMO 2025 and CMO 2025, demonstrating its advanced reasoning and problem-solving capabilities [38][39]. Future Directions - Despite its advancements, DeepSeek acknowledges that V3.2 still has room for improvement in training resource allocation and token efficiency compared to top closed-source models [42][43]. - The company aims to enhance the underlying model and post-training methods in future versions, indicating potential developments for V4 [43].