Core Insights - DeepSeek is advancing towards a "computing power freedom" path with its V3.1 release, optimizing the use of NVIDIA GPU power while adapting to domestic chips, potentially reducing memory usage by up to 75% [4][27]. - The V3.1 upgrade enhances DeepSeek's efficiency in reasoning and tool usage, positioning it competitively against international AI firms [8][9]. Group 1: Technological Advancements - DeepSeek V3.1 introduces a hybrid reasoning architecture, supporting both thinking and non-thinking modes, which improves efficiency and reduces token consumption [6][8]. - The model has undergone extensive retraining with an additional 840 billion tokens, achieving a context length of 128k, which enhances performance while lowering costs [8][9]. - The API Beta interface now supports strict function calling, improving reliability and usability in enterprise applications, making it easier to replace existing solutions like GPT/Claude [9]. Group 2: Market Positioning - DeepSeek's V3.1 is a significant milestone in its transition to the Agent era, allowing for better integration into the enterprise market, particularly with support for Anthropic API formats [9][30]. - The shift towards using UE8M0 FP8 scale data format allows DeepSeek to efficiently run large models on domestic AI chips, reducing reliance on imported GPUs [12][27]. - The potential decline in demand for NVIDIA's H20/B30 chips in China is noted, as domestic chips become more capable of handling large models with the new low-precision training methods [29][30]. Group 3: Competitive Landscape - NVIDIA's long-standing use of low-precision formats has set a benchmark, but DeepSeek's innovations may accelerate the development of domestic chips, creating a more independent AI ecosystem in China [16][32]. - Despite the advancements by DeepSeek, NVIDIA retains advantages in bandwidth, interconnectivity, and a robust software ecosystem, which may still attract international firms [32].