商 品 研 究 2025 年 10 月 22 日 黄金：俄乌危机缓解 白银：现货矛盾缓解，冲高回落 刘雨萱 投资咨询从业资格号：Z0020476 liuyuxuan023982@gtjas.com 【基本面跟踪】 贵金属基本面数据 | 贵金属基本面数据 | | | | | | | --- | --- | --- | --- | --- | --- | | | 沪金2512 | 昨日收盘价 994.06 | 日涨幅 2.45% | 昨日夜盘收盘价 945.44 | 夜盘涨幅 -4.64% | | | 黄金T+D | 986.89 | 1.35% | 944.47 | -4.55% | | | Comex黄金2512 | 4138.50 | -5.39% | - | | | | 伦敦金现货 | #N/A | #N/A | - | - | | | 沪银2512 | 11805 | 0.51% | 11285.00 | -4.86% | | 价 格 | 白银T+D | 11759 | -0.16% | 11261 | -4.86% | | | Comex白银2512 | 48.160 | -6.30% | - | - ...

Core Insights - DeepSeek has introduced a new model called DeepSeek-OCR, which can compress text information into images, achieving a significant reduction in token usage while maintaining high accuracy [5][31][39]. Group 1: Model Capabilities - DeepSeek-OCR can store large amounts of text as images, allowing for a more efficient representation of information compared to traditional text-based models [9][10]. - The model demonstrates a compression ratio where it can use only 100 visual tokens to outperform previous models that required 256 tokens, and it can achieve results with less than 800 visual tokens compared to over 6000 tokens used by other models [14][31]. - DeepSeek-OCR supports various resolutions and compression modes, adapting to different document complexities, with modes ranging from Tiny to Gundam, allowing for dynamic adjustments based on content [17][18]. Group 2: Data Utilization - The model can capture previously unutilized data from documents, such as graphs and images, which traditional models could not interpret effectively [24][26]. - DeepSeek-OCR can generate over 200,000 pages of training data in a day on an A100 GPU, indicating its potential to enhance the training datasets for future models [29]. - By utilizing image memory, the model reduces the computational load significantly, allowing for a more efficient processing of longer conversations without a proportional increase in resource consumption [31]. Group 3: Open Source Collaboration - The development of DeepSeek-OCR is a collaborative effort, integrating various open-source resources, including Huawei's Wukong dataset and Meta's SAM for image feature extraction [38][39]. - The model's architecture reflects a collective achievement from the open-source community, showcasing the potential of collaborative innovation in AI development [39].

Report Summary 1. Report Industry Investment Rating - Not provided in the given content. 2. Core Viewpoints - The global economic situation is complex, with the US economy affected by wrong policies and the global economy entering the top - region [2]. - AI - related fields are rapidly developing, with multiple technological breakthroughs and new product launches, and AI toys are reshaping the consumer market [1]. - The US banking industry has increased concerns about credit losses, leading to increased merger expectations [1]. - The Japanese stock market has reached new highs, but the market is now focusing on the political stability of the new government [1]. 3. Summary by Related Catalogs 3.1 Important Information - OpenAI is facing absolute computing power scarcity, and it has a resource - allocation mechanism for research and application [1]. - DeepSeek has launched a revolutionary OCR model that solves the computing power problem of AI in processing long documents and has high - efficiency data generation capabilities [1][2]. - Yushu (Unitree) has released a four - legged robot training platform and a 180 - cm humanoid robot Unitree H2 with good balance and control capabilities [1][2]. - AI toys are becoming a new consumer hotspot across all age groups and are in a "golden age" of development, breaking the scene boundaries of traditional toys [1]. - The nuclear fusion technology competition is intensifying in the US, covering two key tracks: small modular reactors for AI power and future - oriented fusion technology [1]. - US banks' concerns about credit losses are promoting merger expectations, and regulatory and political factors support merger discussions [1]. - Kōichi Hamada has been elected as the new Prime Minister of Japan, and the Japanese stock market has reached new highs, but market attention has shifted to the political stability of the new government [1]. 3.2 Global Economic Logic - China's industrial added value in September 2025 increased by 6.5% year - on - year and 0.64% month - on - month, and the central parity rate of the RMB against the US dollar has been raised to 7.0995 [1]. - Traders are betting that the Fed will cut interest rates by at least 50 basis points in the upcoming meetings [1]. - China is the preferred stock investment market for emerging markets, and 100 surveyed institutions manage $423 billion in emerging - market assets [2]. - Huawei has announced the evolution and goals of Ascend chips, with its computing power "super - node + cluster" leading NVIDIA by more than a year [2]. - Although AI infrastructure investment has reached a new high in nominal terms, the proportion of US AI investment in GDP is still less than 1% [2]. - TSMC's CEO said that AI demand is stronger than expected [2].

1. Report Industry Investment Rating - No information regarding the industry investment rating is provided in the report. 2. Core Viewpoints - The stabilization of the stock market is crucial for expanding the property - income channels of urban and rural residents and enhancing people's consumption confidence. A stable stock market can inject capital into the real economy and promote consumption through wealth, psychological, and expected effects [1][2][3]. - From the 15th Five - Year Plan period, people's consumption demand will shift from survival - type to development - type, and the consumption structure will change from being dominated by commodity consumption to a balance between commodity and service consumption. Service consumption will be a key area for expanding consumption in China [2]. - The ETF market has seen significant net inflows in October 2025, with equity - based ETFs being the main driving force. Many foreign institutions believe that the current valuation of A - shares is reasonably low, making them attractive for investment [1][3]. 3. Summary by Relevant Catalogs Market Review - On Tuesday, with the rise of overseas markets, the major domestic stock indices fluctuated upwards, and the communication sector led the gains. The total trading volume of the two markets was 1.87 trillion yuan, showing a slight increase. The CSI 300 index closed at 4607 points, up 69 points or 1.53%; the SSE 50 index closed at 3007 points, up 32 points or 1.09%; the CSI 500 index closed at 7185 points, up 115 points or 1.64%; the CSI 1000 index closed at 7344 points, up 104 points or 1.45% [1]. - Among industry and theme ETFs, those related to communication, 5G, and consumer electronics led the gains, while coal, energy, and dividend ETFs led the losses. Among sector indices, consumer electronics, communication equipment, and other sectors led the gains, while forestry, coal mining, and other sectors led the losses [1]. - The futures of the CSI 500, CSI 1000, CSI 300, and SSE 50 indices saw net inflows of 3600 million, 2400 million, 1700 million, and 500 million yuan respectively [1]. Important Information - As of October 13, the number of new A - share accounts exceeded 20 million, a year - on - year increase of over 50%, which has effectively increased residents' property income [1][3]. - Most domestic families allocate over 20% of their financial assets to the securities market, and the fluctuation of stock book value affects residents' wealth and consumption willingness [1][3]. - Morgan Stanley believes that factors such as upcoming dividend distributions, stable interest rates, and 500 billion yuan in structural financial policy tools will support the re - evaluation of Chinese bank stocks [2]. - AI toys are reshaping the industry landscape in the 2025 consumer market, becoming a new consumption hotspot across all age groups [2]. - OpenAI is facing a shortage of computing power, which restricts the release of many products [2]. - Concerns about credit losses in the US banking industry have increased the expectation of mergers and acquisitions [2]. - After the election of the new Japanese Prime Minister, the Japanese stock market has reached new highs, but the market is now focusing on the political stability of the new government [2]. - Morgan Chase believes that the competition in nuclear fusion technology is intensifying, with two key tracks attracting large amounts of capital [2]. Market Logic - With the rise of overseas markets, the domestic stock market rose on Tuesday. The increase in new A - share accounts and the net inflow of funds into the ETF market reflect the growing confidence of investors. The reasonable and low valuation of A - shares makes them attractive to foreign investors [1][3]. Future Outlook - The stock market is expected to remain in a volatile state, and the market is waiting for the clarity of Sino - US negotiations at the end of the month. The long positions of stock index futures should be mainly based on the CSI 300 and SSE 50 indices [3]. - Traders are increasing their bets on the Fed to cut interest rates by at least 50 basis points in the upcoming meetings [3]. Trading Strategies - For stock index futures directional trading, due to the volatile market and the uncertainty of Sino - US negotiations, long positions should be mainly based on the CSI 300 and SSE 50 indices [3]. - For stock index option trading, as the market is in a volatile and consolidating state, it is advisable to wait and see [3].

Core Insights - DeepSeek has open-sourced a new model called DeepSeek-OCR, which utilizes visual patterns for context compression, aiming to reduce computational costs associated with large models [1][3][6] Model Architecture - DeepSeek-OCR consists of two main components: DeepEncoder, a visual encoder designed for high compression and high-resolution document processing, and DeepSeek3B-MoE, a lightweight language decoder [3][4] - The DeepEncoder integrates two established visual model architectures: SAM (Segment Anything Model) for local detail processing and CLIP (Contrastive Language–Image Pre-training) for capturing global knowledge [4][6] Performance and Capabilities - The model demonstrates strong "deep parsing" abilities, capable of recognizing complex visual elements such as charts and chemical formulas, thus expanding its application in fields like finance, research, and education [6][7] - Experimental results indicate that when the number of text tokens is within ten times that of visual tokens (compression ratio <10×), the model achieves 97% OCR accuracy, maintaining around 60% accuracy even at a 20× compression ratio [6][7][8] Industry Reception - The model has received widespread acclaim from tech media and industry experts, with notable figures like Andrej Karpathy praising its innovative approach to using pixels as input for large language models [3][4] - Elon Musk commented on the long-term potential of AI models primarily utilizing photon-based inputs, indicating a shift in how data may be processed in the future [4] Practical Applications - DeepSeek-OCR is positioned as a highly practical model capable of generating large-scale pre-training data, with a single A100-40G GPU able to produce over 200,000 pages of training data daily [7][8] - The model's unique approach allows it to compress a 1000-word article into just 100 visual tokens, showcasing its efficiency in processing and recognizing text [8]

Liquidmetal Technologies Conference Summary Company Overview - **Company**: Liquidmetal Technologies (OTCPK:LQMT) - **Founded**: 1987, with technology originating from Caltech in 1962 - **IPO**: 2002 on NASDAQ with initial orders from Samsung for flip phone hinges [2][3] - **Current Focus**: Manufacturing and commercialization of liquid metal technology, particularly for hinges and other applications in various industries [1][9] Key Technology Insights - **Liquid Metal Technology**: Utilizes amorphous alloys, primarily a zirconium-based alloy, which is 70% zirconium and includes titanium, nickel, and aluminum [4][5] - **Manufacturing Process**: Involves a hybrid die-cast injection molding machine, allowing for the production of parts that are stronger than titanium and have superior hardness and elasticity [5][6] - **Unique Selling Proposition**: Capable of producing parts that are thinner (0.3 mm) and lighter, making them ideal for modern mobile devices [6][7] Target Industries - **Medical Devices**: High potential for complex, high-tolerance parts such as surgical tools and pacemaker housings [7][15] - **Robotics and Electric Vehicles (EVs)**: Applications in robotics (e.g., Tesla Optimus) and EV components, including parts for Tesla Model X [8][15] - **Consumer Products**: Prototyping for various consumer items, including health rings, credit cards, earbuds, and sunglasses [8][15] Future Growth and Manufacturing Plans - **New Manufacturing Plant**: Set to open in Hangzhou, China in 2026, leveraging local innovation and manufacturing expertise [9][10] - **Chairman’s Role**: Professor Lugee Li, who invested $63 million in 2016, is leading the operations and has a strong background in manufacturing [10][11] Market Potential - **Foldable Devices Market**: Estimated to grow from $1 billion in 2024 to $7 billion in 10 years, with significant revenue potential from hinge production [12][13] - **Revenue Opportunities**: Potential to manufacture millions of parts, translating to substantial revenue from single applications [12][13] Competitive Landscape - **Main Competitors**: CNC machining and metal injection molding (MIM) processes, with Liquidmetal's technology being more cost-effective and precise [19][20] - **Cost Structure**: Parts priced between $1 to $10, depending on complexity and production volume, making them competitive against traditional manufacturing methods [19][20] Intellectual Property and Market Position - **Patents**: Approximately 40 patents held, with plans to focus on developing additional patents to protect technology [22] - **Market Leadership**: Positioned as the foremost authority in amorphous alloy technology, with a strong brand recognition compared to smaller players in China [15][22] Financial Health - **Current Stock Price**: Ranges from $0.13 to $0.15, with a healthy balance sheet showing about $40 million in liquid cash and assets [17] - **Future Plans**: Aiming for potential re-listing on NASDAQ by 2026, with no immediate plans to raise additional funds [17][18] Conclusion - **Outlook**: The future appears bright for Liquidmetal Technologies, with numerous revenue opportunities and a strong focus on innovation and market expansion in various high-demand industries [18][23]

Group 1 - Anthropic has launched the web version of Claude Code, allowing users to delegate programming tasks directly from the browser, with tasks running on cloud infrastructure [1] - The Claude Code feature supports parallel execution of multiple programming tasks and can connect to GitHub repositories to automatically create pull requests [1] - The iOS app has also synchronized the Claude Code feature, enabling developers to program anytime and anywhere, particularly useful for handling backlog issues and routine fixes [1] Group 2 - Tsinghua University and Zhizhu have jointly launched the Glyph framework, which renders text information into images for processing with visual models, achieving a text compression rate of 3-4 times [2] - Glyph employs a three-stage method of continuous pre-training, LLM-driven rendering search, and post-training, using genetic algorithms to find optimal rendering configurations [2] - Glyph complements the DeepSeek-OCR path, with DeepSeek extracting information from images to validate the feasibility of visual compression, while Glyph verifies contextual expansion capabilities by converting text to images [2] Group 3 - Elon Musk announced that the X platform will completely remove heuristic recommendation algorithms in favor of Grok, which will automatically match user interests by reading and watching all content [3] - Heuristic algorithms rely on human-set rules, leading to dominance by large accounts and lack of exposure for quality content from new accounts; Grok will allow for fairer content distribution [3] - Users can dynamically adjust content recommendations with Grok, sparking discussions about the "death of the internet" theory, suggesting AI is ending the essence of human interaction in social media [3] Group 4 - Adobe has launched the AI Foundry service, allowing businesses to collaborate with Adobe to build proprietary generative AI models based on their own brand and intellectual property [4] - The service is supported by the Firefly series of models, which are trained using fully licensed data, and operates on a pay-per-use basis [4] - Since the launch of Firefly, businesses have generated over 25 billion creative assets, with future integration into Microsoft core products like Copilot and Bing Image Creator [4] Group 5 - Sogou Input Method has introduced the first AI companion assistant for computers, "Xiao Wan," based on Tencent's mixed Yuan model, providing emotional support and companionship in the workplace [6] - Tencent Video has launched an exclusive AI companion for the drama "Allow Me to Shine," featuring a character-based AI that engages in realistic conversations through text and voice [6] - The mixed Yuan AI companion is capable of understanding dialogue context, multi-turn conversations, and tool invocation, enhancing character role-play through deep training [6] Group 6 - McKinsey received a token consumption award from OpenAI, indicating significant spending on strategic consulting presentations that were largely generated by ChatGPT [7] - Since launching its internal AI Lilli in 2023, over 70% of McKinsey's 40,000 employees use the platform, which responds to over 500,000 queries monthly, despite a workforce reduction of over 5,000 employees [7] - AI startups like PromptQL and Parable AI are capturing market share from second-tier consulting firms, leading to a 54% year-on-year drop in entry-level job postings in the consulting industry [7] Group 7 - Anthropic has launched Claude for Life Sciences, a specialized version of Claude designed for life sciences, achieving a score of 0.83 on the Protocol QA benchmark, surpassing the human benchmark of 0.79 [8] - The new version includes connectors for various research platforms, supporting large-scale bioinformatics analysis [8] - It offers specialized skills for literature reviews, experimental design, bioinformatics analysis, and regulatory compliance, covering the entire process from early discovery to results translation [8] Group 8 - DeepSeek has released the open-source model DeepSeek-OCR, which proposes a "contextual optical compression" approach, achieving a compression rate of 10 times with an OCR decoding accuracy of 97% [9] - The model utilizes a DeepEncoder and DeepSeek3B-MoE-A570M architecture, supporting various input modes and achieving new state-of-the-art results on OmniDocBench [9] - The research introduces the idea of simulating human memory mechanisms through optical compression, providing new directions for constructing infinitely long contextual architectures [9] Group 9 - Jason Wei, a former core researcher at OpenAI, outlined three key ideas for understanding AI development in 2025: the verifier's law, the commodification of intelligence, and the jagged edge of intelligence [10] - The verifier's law includes five dimensions of verifiability: objectivity, verification speed, batch verifiability, low noise, and continuous feedback, suggesting that any task that is solvable and easily verifiable will eventually be tackled by AI [10] - The most significant impact of AI will be in digital tasks that are not difficult for humans and are data-rich, with areas like software development seeing accelerated progress, while non-digital tasks will remain unchanged [10]

以下文章来源于AI产业链研究 ，作者研究 AI产业链研究 . 围绕人工智能展开研究，涵盖基础设施、算法及应用等多个方面，同时也会分享研究过程中的一些心得 体会 中国开源模型在海外逐渐占据越来越大的市场份额是不争的事实。关于中国开源模型的讨论也越来越 多，DeepSeek 本周新推出的一款 OCR 模型更是在X上引发广泛关注 —— 这实际上是一款新发布的开 源视觉语言模型（VLM）。 它并不是又一款普通的 OCR 工具，而是 "光学上下文压缩" 领域的突破性成果：将图像作为编码和处理 海量文本数据的超高效率载体，成功解决了大型语言模型（LLMs）的核心痛点之一 —— 在处理长上下 文时，避免内存、延迟或令牌成本的激增。当然，更关键的是它的开源属性。 这几天，一张图片在海外 AI 圈刷屏—— DeepSeek 在投资领域的表现同样亮眼。在 2025 年 10 月的 Alpha Arena 赛事（Hyperliquid 平台举办的实盘加密货币交易竞赛）中，DeepSeek-V3.1 以 1 万美元本 金参赛，三天内斩获 40.4% 的回报率登顶排行榜 —— 不仅超越了 Grok 4（33.4%）和 Claude （ ...

Our story begins in 2023 in Hjo, China. A team of hedge fund analysts made a bet over lunch. They spent years building algorithms predict to predict stock markets.But today they're arguing about something different. Could they make an AI chatbot that can potentially rival Chad PG. Not for profit, not for glory, but for fun.Fast forward 18 months. That chatbot deep seats now outperforms Chhat's Chhat's premium models in both programming and math and is completely free and is rewriting the rule book for what ...

Core Insights - DeepSeek has open-sourced DeepSeek-OCR, an OCR model that achieves state-of-the-art results on benchmarks like OmniDocBench [1] - The motivation behind entering the OCR field is to address the computational bottleneck of long context processing in large language models (LLMs) [4][6] - The paper proposes that text information can be efficiently compressed through optical 2D mapping, allowing visual language models (VLMs) to decompress original information from images [4][6] Group 1: Long Context Processing - The pursuit of longer context in LLMs has led to a competitive arms race, with token windows expanding from thousands to millions [7] - The core limitation arises from the attention mechanism in the Transformer architecture, where computational complexity and memory usage grow quadratically with sequence length [7] - DeepSeek-AI's engineers propose a fundamental question: can the number of tokens be compressed rather than just optimizing attention calculations? [7][10] Group 2: Visual Tokens vs. Text Tokens - Visual tokens are the basic units of information processed by visual models, while text tokens are used by LLMs [8] - A 1024x1024 image can be divided into 4096 visual tokens, significantly reducing the number of tokens needed compared to text representation [9] - The understanding that visual modalities can serve as efficient compression mediums for text information led to the creation of DeepSeek-OCR [9] Group 3: DeepEncoder and Compression Techniques - DeepSeek-OCR is essentially a proof of concept for an "optical compression-decompression" system [10] - The DeepEncoder, a key innovation, is designed to handle high-resolution inputs while producing minimal visual tokens [11][12] - The architecture consists of three stages: a local detail processor, a compression module, and a global attention layer [14][16] Group 4: Performance Metrics - Experimental results show a 10.5x compression rate with 64 visual tokens decoding 600-700 text tokens, achieving an OCR accuracy of 96.5% [17][18] - At a 20x compression rate, the model maintains around 60% accuracy while decoding over 1200 text tokens [17][18] - DeepSeek-OCR outperforms existing models like GOT-OCR2.0 and MinerU2.0 in terms of performance and token efficiency [19][20] Group 5: Future Vision and Memory Simulation - The team aims to simulate human memory's forgetting mechanism, which naturally prioritizes relevant information while compressing less important details [25][27] - The multi-resolution design of DeepSeek-OCR provides a technical foundation for managing memory in a way that mimics human cognitive processes [29][30] - The ultimate goal is to create a system that balances information retention and computational efficiency, potentially leading to a new paradigm in AI memory and input systems [32][35]