思维链压缩训练

Search documents
DeepSeek-V3.1发布:更高效思考、更强Agent能力、更长上下文
生物世界· 2025-08-21 08:00
Core Insights - DeepSeek has officially released DeepSeek-V3.1, introducing a hybrid reasoning architecture that allows users to switch between "Deep Thinking" mode and "Non-Thinking" mode for enhanced interaction [2][3]. Group 1: Hybrid Reasoning Architecture - The "Deep Thinking" mode (DeepSeek-Reasoner) is designed for tasks requiring deep reasoning, such as mathematical calculations and complex logic analysis, providing higher reasoning efficiency [3]. - The "Non-Thinking" mode (DeepSeek-Chat) is tailored for everyday conversations and information queries, offering quicker responses [4]. - Users can easily switch modes via a "Deep Thinking" button on the official app and web interface, enhancing the user experience [5]. Group 2: Enhanced Agent Capabilities - DeepSeek-V3.1 has significantly improved tool usage and agent task performance through Post-Training optimization, resulting in fewer required iterations and higher efficiency in code repair and command line tasks [6]. - Benchmark results show that DeepSeek-V3.1 outperforms its predecessor, DeepSeek-R1-0528, in various tasks, including SWE-bench and Terminal-Bench, with scores of 66.0 and 31.3 respectively [7][8]. Group 3: Efficiency Improvements - The new version employs a thought chain compression training method, reducing output tokens by 20%-50% while maintaining performance levels comparable to DeepSeek-R1-0528, leading to faster response times and lower API call costs [9]. Group 4: API Upgrades and Model Availability - The DeepSeek API has been upgraded to support a context length of 128K, facilitating easier handling of long documents [10][12]. - The base and post-training models of DeepSeek-V3.1 are now open-sourced on platforms like Hugging Face and ModelScope, with a price adjustment for the API set to take effect on September 6, 2025 [11].