DeepSeek正在改写AI的扩展法则：下一代AI的决胜点不再是单纯堆砌更大的GPU集群，而是通过更聪明的混合架构，用性价比更高的DRAM置换 稀缺的HBM资源。 据追风交易台消息，摩根士丹利1月21日发布的最新研报显示，DeepSeek正在通过一种名为"Engram"的创新模块，改变大语言模型的构建方式。 其核心突破在于将存储与计算分离，通过引入"条件记忆"（Conditional Memory）机制，大幅减少了对昂贵且紧缺的高带宽内存（HBM）的需 求，转而利用成本更低的普通系统内存（DRAM）来处理复杂的推理任务。 大摩分析师Shawn Kim及其团队认为，DeepSeek展示了如何"少花钱多办事"（Doing More With Less）的哲学。这种将存储与计算分离的技术路 径，不仅缓解了中国面临的AI算力约束，更向市场证明了高效的混合架构才是AI的下一个前沿。 这一被大摩重点关注的架构，源自DeepSeek创始人梁文锋团队与北大合作者在1月13日发布的重磅论文《Conditional Memory via Scalable Lookup》。在这篇论文中，团队首次提出了"Engram"（印迹）模块。 ...

技术表现方面，新模型端到端 S2S 架构可以从语音输入直接生成语音输出，无需 ASR + LLM + TTS 多 模块拼接，效率更高、延迟更低；Shared LLM 层以 5Hz 帧率 高效处理，SRH 以 25Hz 帧率 生成高质量 语音，GPU 计算开销降低近 50%；训练内容覆盖音频理解、语音问答、情感识别、工具调用等真实场 景，让模型更"接地气"。 据称，该模型不是简单的"能聊天"，而是听得懂你的话、感知你的情绪、还能帮你真正干活的AI语音搭 子。 北京商报讯（记者 陶凤 王天逸）12月23日，通义大模型官方发布了新一代端到端语音交互模型 Fun- Audio-Chat。 ...

Core Viewpoint - The article emphasizes the potential of investing in the AI industry through ETFs, particularly the Vanguard Information Technology ETF, which provides diversified exposure to leading technology companies involved in AI [1][3]. Industry Overview - The global AI market is projected to exceed $826 billion by 2030, indicating significant growth potential despite its unpredictability [1]. - Advancements in AI could lead to developments such as humanoid robotics, transitioning from science fiction to reality [2]. ETF Analysis - The Vanguard Information Technology ETF (VGT) is highlighted as a suitable investment for those seeking growth without the complexities of selecting individual AI stocks [3]. - Although not a dedicated AI ETF, VGT includes many leading AI companies among its top holdings, such as Nvidia, Apple, and Microsoft, which are integral to the AI ecosystem [4][5]. - The ETF's top 10 holdings include major players in the technology sector, reinforcing its relevance to the AI market [6]. Performance Metrics - VGT has a long-standing track record of outperforming the broader stock market, attributed to the increasing importance of technology in the economy [9]. - The ETF charges a low expense ratio of 0.09%, which is significantly lower than many dedicated AI ETFs, potentially enhancing long-term investment returns [8]. Market Dynamics - The technology sector, including AI, is becoming increasingly vital across various industries, with traditional sectors adopting technology for efficiency and optimization [10]. - Despite the potential for explosive growth, the ETF and technology stocks are subject to volatility, with historical declines noted during market downturns [12][13].

Investment Rating - The report does not explicitly provide an investment rating for the AI industry or specific companies within it. Core Insights - The competition between the US and China in the AI sector is intensifying, with both countries making significant strides in technology and investment to secure leadership in AI development [2][4][12] - Chinese companies are rapidly advancing in AI capabilities, with models that are approaching those of leading US firms, driven by government support and a focus on open-source technologies [3][8][9] - The outcome of the AI race may determine the technological superpower of the 21st century, with both nations prioritizing AI for economic, political, and national defense purposes [4][12][13] Summary by Sections The Technology - The US has led key breakthroughs in AI, with companies like OpenAI and Alphabet pioneering advanced computing chips and large language models [6][7] - Chinese firms are quickly following suit, developing AI models that require less computational power and embracing open-source standards to enhance global adoption [8][9] The State - AI is a national priority for both the US and China, with the US aiming to maintain a technological edge and China promoting AI as a public good [12][13] - The US government has initiated plans to reduce regulatory barriers for AI development, while China emphasizes the need for international cooperation in AI governance [12][13] The Money - In the first half of 2025, US AI startups raised over $100 billion, while major tech firms are projected to spend more than $344 billion on AI infrastructure [26][27] - China's AI capital expenditure is expected to reach $98 billion in 2025, a 48% increase from 2024, with significant government backing [27][28] The Talent - The US has historically attracted top AI talent from around the world, but tightening visa policies pose risks to this talent pipeline [29][31] - China is actively working to reverse brain drain by attracting scientists and entrepreneurs educated abroad back to the country [32][33] The Infrastructure - China has built a robust AI ecosystem supported by vast data pools and renewable energy-powered data centers [34][41] - The US faces challenges with aging power grids, while China has significantly increased its energy capacity to support AI development [41][42]

Core Insights - The Qwen team has released two new models, Qwen3-4B-Instruct-2507 and Qwen3-4B-Thinking-2507, which are designed to enhance performance on various tasks, particularly in reasoning and general capabilities [2][3][5]. Model Performance - Qwen3-4B-Thinking-2507 achieved a score of 81.3 in the AIME25 assessment, outperforming competitors like Gemini 2.5 Pro and Claude 4 Opus [4][5][23]. - The new models support a context length of 256k, significantly improving context awareness and understanding [3][17]. Model Specifications - Qwen3-4B-Instruct-2507 is a non-reasoning model that enhances general capabilities and multi-language support, while Qwen3-4B-Thinking-2507 is a reasoning model tailored for expert-level tasks [7][16]. - The 4B parameter size is particularly friendly for edge devices, allowing for deployment on small hardware like Raspberry Pi [2][8][26]. Comparative Analysis - In various tests, Qwen3-4B-Instruct-2507 outperformed smaller closed-source models like GPT-4.1-nano and showed comparable performance to larger models like Qwen3-30B-A3B [13][15]. - The models exhibit significant improvements in areas such as instruction following, logical reasoning, and text generation, with enhanced alignment to user preferences [17][24]. Deployment Recommendations - The Qwen team has provided deployment suggestions for local use, including applications like Ollama and MLX-LM, and recommended using a quantized version for very small devices [27][28]. - For optimal performance, especially in reasoning tasks, it is advised to use a context length greater than 131,072 tokens [29]. Community Engagement - The Qwen team has encouraged community feedback and interaction, with links provided for accessing the new models on platforms like Hugging Face and ModelScope [26][36].