智通财经网·2025-08-27 04:46Core Insights - DeepSeek V3.1 significantly outperforms R1-0528 across multiple metrics, enhancing agent support and showcasing innovative use of UE8M0 FP8 Scale precision in collaboration with domestic AI chips [1] Group 1: Major Upgrades in DeepSeek V3.1 - The release of DeepSeek V3.1 includes three major upgrades: a hybrid reasoning architecture that supports both thinking and non-thinking modes, an official app and web model upgrade, and a "deep thinking" button for mode switching [1] - Enhanced thinking efficiency allows DeepSeek V3.1-Think to provide answers in a shorter time compared to DeepSeek-R1-0528 [1] - Improved agent capabilities through Post-Training optimization, resulting in better performance in tool usage and agent tasks [1] Group 2: Enhanced Tool and Agent Support - Programming agents show significant improvement in code repair assessments and complex tasks in terminal environments compared to previous DeepSeek models [2] - Search agents in DeepSeek V3.1 have achieved substantial enhancements in multiple search evaluation metrics, particularly in complex multi-step reasoning and expert-level problem tests [2] - Thinking efficiency has improved, with V3.1-Think maintaining average performance on par with R1-0528 while reducing output token count by 20%-50% [2] Group 3: API and Model Open Source - The Base model of V3.1 underwent extensive retraining, adding a total of 840 billion tokens, and is now available on Hugging Face and MoDa [3] - DeepSeek V3.1 utilizes UE8M0 FP8 Scale precision, specifically designed for the upcoming generation of domestic chips [3] - Significant adjustments have been made to the tokenizer and chat template, resulting in notable differences from DeepSeek V3 [3]