DeepSeek，重磅发布！

Core Viewpoint - DeepSeek officially released DeepSeek-V3.1 on August 21, showcasing significant upgrades in its capabilities and performance metrics [1]. Summary by Sections Upgrade Features - The official app and web model have been upgraded to DeepSeek-V3.1, allowing users to switch between thinking and non-thinking modes via a "Deep Thinking" button [2]. - The DeepSeek API has also been upgraded, with deepseek-chat corresponding to non-thinking mode and deepseek-reasoner to thinking mode, expanding context to 128K [2]. - The API Beta interface supports strict mode Function Calling to ensure outputs meet schema definitions [2]. Performance Improvements - DeepSeek-V3.1 has shown substantial improvements in multiple search evaluation metrics, particularly in complex search tests requiring multi-step reasoning (browsecomp) and expert-level multidisciplinary problem tests (HLE), outperforming R1-0528 significantly [2]. - The model utilizes UE8M0 FP8 Scale for parameter precision, and there have been major adjustments to the tokenizer and chat template, resulting in noticeable differences from DeepSeek-V3 [2]. Market Reaction - DeepSeek concept stocks experienced a sharp rise in the market towards the end of the trading session [3]. Architectural Enhancements - The new mixed reasoning architecture supports both thinking and non-thinking modes within a single model [5]. - Enhanced thinking efficiency allows DeepSeek-V3.1-Think to provide answers in a shorter time compared to DeepSeek-R1-0528 [5]. - The agent capabilities have improved significantly through post-training optimization, enhancing performance in tool usage and intelligent agent tasks [5].