集成电路（芯片）： 1. 理想汽车：理想L6累计交付突破240,000辆。 2. 禾赛科技与零跑汽车达成合作 正式官宣20万台激光雷达订单。 3. 可自定义添加功能，小米汽车支持苹果Siri语音控车。 4. 长安汽车4月新品计划公布，含阿维塔06/07探索版、启源Q07等9款车型。 5. 比亚迪：预计一季度盈利人民币85亿元–100亿元，同比上升86.04%-118.88%。 1. 我国二维半导体芯片研发取得重要突破。 2. 美光告诉美国客户，将对部分产品征收与关税相关的附加费。 3. 瑞芯微：市场上有谣言声称"三星晶圆厂暂停所有中国业务"。经与三星晶圆厂证实，此为假消息。 其他： 1. 酷狗音乐与DeepSeek达成深度合作。 2. 白宫：特朗普相信美国有能力制造iPhone。 3. 开源模型DeepCoder媲美OpenAI-o3。 4. 广州市番禺区发布全市首个"空中的士"快速出行网络。 5. 工业富联：一季度净利润同比预增24.4%—26.8%。 6. 长城汽车：与宇树科技签署战略合作协议，将在机器人技术等领域合作。 金十数据整理：每日科技要闻速递（4月9日） 新能源汽车： ...

出品 | 虎嗅科技组 作者 | 房晓楠 编辑 | 苗正卿 头图 | 视觉中国 无疑，这几天AI圈最大的流量都被Meta承包了。 先是4月6日，迟迟不出手的Meta终于亮相，一举扔出包括Llama 4 Scout（109B）、Llama 4 Maverick （400B）和Llama 4 Behemoth在内的Llama 4套餐，凭借"原生多模态MoE架构""性能超越DeepSeek V3""1000万token上下文"等亮点，一路高举高打。业界也是喜闻乐见，"开源之光依旧能打"等声音持续 不断，Llama 4出道即巅峰。 但没想到，紧接着负面声音层出不穷。 一方面，开发者在实际测评时发现，Llama 4的性能并没有官方宣扬的那样强大，甚至在代码、逻辑推 理方面，远不如GPT-4o、DeepSeek R1 、 Gemini 2.5 pro。 另一方面，有自称Meta内部员工的人员爆料，Llama 4存在造假嫌疑，为了"赶工期"，在后训练阶段 中，将多个benchmark测试集混入训练数据，以提升基准分数。甚至，技术负责人看不过去这样的造假 行为，递交辞职报告。传闻一出，立即发酵，Meta陷入舆论风波中，各种声讨 ...

Core Viewpoint - The company "月之暗面" has announced a significant price reduction for its model inference services and context caching, aiming to remain competitive in the AI model market amidst rising competition from "DeepSeek" [2][3][5]. Pricing Adjustments - The price for model inference services has been reduced from 12 RMB/M Tokens to 2 RMB/M Tokens for input and from 10 RMB/M Tokens to 10 RMB/M Tokens for output for the 8k context length model. For the 32k context length, prices dropped from 24 RMB/M Tokens to 5 RMB/M Tokens for input and from 20 RMB/M Tokens to 20 RMB/M Tokens for output. The 128k version saw a reduction from 60 RMB/M Tokens to 10 RMB/M Tokens for input and from 30 RMB/M Tokens to 30 RMB/M Tokens for output [3][4]. - Context caching prices have also been adjusted, with Cache creation costs dropping from 24 RMB/M Tokens to 4 RMB/M Tokens, Cache storage from 5 RMB/M Tokens/Minute to 1 RMB/M Tokens/Minute, and Cache calls from 0.02 RMB/Request to 0.01 RMB/Request [3][4]. Competitive Landscape - The price reduction represents a move into the competitive landscape of AI models, where "月之暗面" had previously been less aggressive. The company is now joining the price war initiated by "DeepSeek" [5][6]. - Industry experts believe that the timing of this price cut is late, as the first half of the year is critical for AI startups, and competitors are aggressively using price reductions to capture market share [5][6]. User Engagement and Market Position - "月之暗面" experienced significant user engagement, with Kimi's monthly visits reaching 12.18 million in March 2024, making it the second most visited AI assistant after Baidu's Wenxin Yiyan [6]. - However, recent data shows a decline in its market position, with "DeepSeek" leading in active user numbers, significantly impacting "月之暗面" [6][7]. Technological Innovations - The company has attributed the price reductions to technological innovations, particularly through its collaboration with Tsinghua University on the Mooncake project, which has improved inference speed and reduced costs [4][8]. - Despite the advancements, "月之暗面" faces challenges in commercializing its technology effectively, as it lacks the capital backing that competitors like "DeepSeek" possess [7][8]. Future Directions - The company has announced plans to open-source some of its achievements, although experts suggest that technological superiority will be the key factor in determining market success rather than the choice between open-source and closed-source models [8].

Core Viewpoint - The article discusses the shift towards open-source models in the AI industry, highlighting that 2025 marks a significant turning point as major tech companies embrace open-source strategies despite the initial success of closed-source models in commercialization [2][3]. Group 1: Open-source vs Closed-source - The "closed-source" camp focuses on monetization through technology protection, ensuring service quality and data security, while the "open-source" camp promotes accessibility and innovation through shared models and community collaboration [3]. - The rise of open-source models, exemplified by companies like DeepSeek, has initiated an unprecedented "open-source wave" in the global AI industry [3]. Group 2: Major Players and Their Contributions - Major tech companies have released numerous open-source models, with significant contributions from firms like OpenAI, Google, Meta, and Alibaba, showcasing advancements in model performance and capabilities [2][5][6]. - Notable releases include Meta's Llama 4, which is highlighted as one of the most advanced multi-modal models, and DeepSeek's models that have achieved top rankings in open-source performance [5][6]. Group 3: Drivers of Open-source Adoption - The article identifies four key drivers behind the surge in open-source models: the rise of edge intelligence, the need for industry-specific customization, accelerated ecological division of labor, and the crossing of a technological threshold that enhances model usability [11][12][13]. - Open-source models are seen as a means to democratize technology, reduce costs, and foster innovation among developers and small enterprises [14][15]. Group 4: Commercialization Strategies - Companies are exploring various commercialization strategies for open-source models, including offering basic models for free while charging for premium API services, creating community and enterprise versions, and leveraging cloud platforms for monetization [16][17][20]. - The trend indicates a move towards hybrid models that balance open-source initiatives with sustainable revenue generation [20].

初步分析表明，模型回复风格与语气是重要影响因素 （详见风格控制排名） ，我们正在进行更深入的分析！ （比如表情符号控制？） 此外，我们即将在Arena平台 上线Llama-4-Maverick的HuggingFace版本，排行榜结果将稍后公布 。 Meta对我们平台政策的理解与我们对模型提供商的期待存在偏差——Meta本应明确标注 "Llama-4-Maverick-03-26- Experimental"是经过人类偏好优化的定制模型 。 为此，我们正在更新排行榜政策，以强化对公平性、可复现性评估的承诺，避免未来再出现此类混淆。 克雷西 西风 发自 凹非寺 量子位 | 公众号 QbitAI Llama 4真要被锤爆了，这次是大模型竞技场 （Chatbot Arena） 官方亲自下场开怼： 竞技场上，Meta提供给他们的是特供版！ 以下是竞技场背后lmarena.ai团队的原话： 我们注意到社区对Llama-4最新版本在Arena平台的发布存在疑问。为确保完全透明，现 公开2000余组模型对战数据供公众审阅 ，包 含用户提示词、模型回复及用户偏好数据 （链接详见下一条推文） 。 总结一下就是： 官方下场表态后， ...

Group 1 - President Trump enacted a two-step tariff strategy starting April 5, imposing a baseline tariff of 10% on imports from various countries [1] - The stock market reacted negatively, particularly the Nasdaq Composite, which fell 5.8% on April 4 and was down 22% from its December record, entering a bear market [2][3][10] - Major tech stocks like Apple, NVIDIA, and Tesla experienced significant declines due to their exposure to China and the impact of retaliatory tariffs [4][12] Group 2 - Concerns are rising that the investment boom in AI infrastructure is outpacing actual demand, with Alibaba's co-founder warning about oversupply [7] - Microsoft has canceled certain data center projects despite earmarking $80 billion for expansion in 2024, indicating potential oversupply issues [7] - Despite bearish sentiment, major tech companies are committed to over $300 billion in capital expenditures, suggesting potential buying opportunities [8] Group 3 - The Nasdaq 100's price-to-earnings (P/E) ratio has declined from 41.24X in early September 2024 to 29.27X at the end of March 2025, indicating valuation corrections [9][11] - The Nasdaq-100-based ETF Invesco QQQ Trust shows a bullish signal as the 50-day moving average has risen above the 200-day moving average [13] - Investors with a strong risk appetite may consider Nasdaq-100-based ETFs like Invesco QQQ Trust, which currently holds a Zacks Rank 3 (Hold) [14]

Investment Rating - The report does not explicitly provide an investment rating for the industry Core Insights - China's venture capital (VC) activity has significantly weakened amid economic uncertainty and escalating tech rivalry with the US, leading to a decline in private investment across key deep-tech sectors. In response, a state-backed high-tech fund was introduced in March 2025, aiming to inject 1 trillion yuan (approximately $138 billion) over 20 years to revitalize innovation and strengthen the high-tech VC ecosystem in China [3][4][5] - The fund will allocate capital across various sectors, with expected positive impacts in AI & ML and semiconductors, although there are concerns about potential crowding out of private capital. In contrast, sectors like quantum computing and hydrogen energy remain heavily state-driven, limiting VC participation [3][4][5] - The success of the high-tech fund hinges on its ability to integrate private capital effectively. If structured well, it could stabilize the deep-tech VC ecosystem; however, if dominated by state-backed firms, private investors may remain cautious, limiting the fund's long-term impact on innovation and market growth [3][5][58] Summary by Sections Key Takeaways - China's VC activity has weakened, prompting the introduction of a high-tech fund to inject 1 trillion yuan over 20 years to counteract declining private capital flows [3] - The fund's impact will vary across sectors, with potential revitalization in AI & ML but risks of crowding out in semiconductors [3][4] - Past state investment programs indicate that government-linked firms may be favored, posing risks of capital misallocation and restricted private market participation [3][5] China's Strategic Push for High-Tech Leadership - The high-tech fund aims to accelerate innovation in critical industries, including AI, quantum computing, semiconductors, and clean energy, amidst geopolitical tensions and economic challenges [4][7] - The fund's introduction reflects China's urgency to achieve technological self-sufficiency and maintain strategic autonomy in high-tech sectors [4][8] More About the Fund and Its Impact on Private Markets - The high-tech fund will deploy capital across multiple strategic sectors, unlike previous state investment programs that focused on single industries [10] - The fund's structure and execution will determine whether it creates new opportunities for VC and PE investors or follows past patterns of government-driven capital misallocation [5][10] Comparing the High-Tech Fund with the Big Fund - The high-tech fund is designed to support a broader range of industries compared to the Big Fund, which primarily focused on semiconductors [50][52] - The Big Fund's experience highlights risks such as crowding out of private investment and capital misallocation, which could also emerge with the high-tech fund if it favors SOEs [50][58] Concluding Thoughts - The high-tech fund's success will depend on effective capital allocation and the ability to attract private investment, as the current VC landscape in China faces significant challenges [61][64] - Without clear governance and market-based incentives, there is a risk that the fund may replicate past state-led initiatives, limiting private investor participation [64]

让美国Open AI的"o1"和中国DeepSeek的"R1"解答了2025年度东京大学入学考试题目。在被视 为日本国内最难考的东京大学理科3类入学考试中，二者均超过了最低合格线。各科目的具体分数 为…… 在数学方面，虽然很多题目的最终答案正确，但在图形和论证问题上多次出现论述错误和说明不足的情 况。在理科数学方面，在满分120分的情况下，"o1"得了38分，"R1"得了49分。负责数学教学的讲师香 坂季京指出："这被认为明显低于合格者的平均分"。 版权声明：日本经济新闻社版权所有，未经授权不得转载或部分复制，违者必究。 日经中文网 https://cn.nikkei.com 样，包括文科在内，都获得了"合格"成绩。 在二次考试中得分率较高的科目是英语，"o1"和"R1"的得分率都超过了75％。在河合塾负责英语教学的 讲师久恒秀雄表示："单词和语法几乎没有错误。远远超过了参加东京大学入学考试的考生的平均水 平"。 日本经济新闻（中文版：日经中文网）与民营企业和大型预备学校的联合调查显示，中美的新型生成式 AI已经具备了可通过被视为日本国内最难考的东京大学理科3类入学考试的"学力"。调查同时发现，虽 然这些AI ...

Investment Rating - The report does not explicitly state an investment rating for the industry, but it highlights the potential for growth in the domestic AI model sector and suggests focusing on the domestic computing power industry chain [3]. Core Insights - The domestic AI model is accelerating its iteration, with new models like Doubao and AutoGLM demonstrating superior performance and cost efficiency compared to existing models like DeepSeek. Doubao's approach of "deep thinking first, then online search" is noted to yield better accuracy and completeness [3]. - The report emphasizes the shift in AI applications from simple chatbot forms to more complex AI agents, indicating a significant increase in computational power requirements as these products are developed [3]. - The report anticipates a substantial growth in generative AI spending, projected to reach $644 billion in 2025, driven primarily by hardware investments [18]. Weekly Market Review - The TMT sector experienced declines across major industries: Communication (-0.79%), Media (-1.28%), Computer (-1.87%), and Electronics (-2.71%). Notably, the education publishing sector saw a gain of 3.38% [3][11]. - The top-performing stocks in the electronics sector included ST Yushun (21.54%), Xuguang Electronics (16.17%), and Changyang Technology (14.51%) [10][12]. Industry News and Key Company Announcements - Significant events include the launch of new AI models and conferences focused on AI and technology, such as the AIoT conference and the Global 6G Technology and Industry Ecosystem Conference [16]. - The report mentions that DeepSeek has become one of the fastest-growing AI tools globally, with a monthly visitor count surpassing that of ChatGPT [17]. - Huawei's annual report indicates a 22.4% increase in global sales revenue, reaching 862.1 billion RMB, while R&D investment was 179.7 billion RMB, accounting for 20.8% of total revenue [22].

Core Viewpoint - Meta has launched the Llama 4 series, which includes the most advanced models to date, Llama 4 Scout and Llama 4 Maverick, marking a significant advancement in open-source AI models and a response to emerging competitors like DeepSeek [1][3][10]. Group 1: Model Features - Llama 4 series includes two efficient models: Llama 4 Scout and Llama 4 Maverick, with a preview of the powerful Llama 4 Behemoth [5][8]. - The Llama 4 models utilize a mixture of experts (MoE) architecture, enhancing computational efficiency by activating only a small portion of parameters for each token [7][8]. - Llama 4 Behemoth boasts a total parameter count of 2 trillion, while Llama 4 Scout has 109 billion parameters and Llama 4 Maverick has 400 billion parameters [8]. Group 2: Multi-Modal Capabilities - Llama 4 is designed as a native multi-modal model, employing early fusion technology to integrate text, images, and video data seamlessly [8][9]. - The model supports extensive visual understanding, capable of processing up to 48 images during pre-training and 8 images during post-training, achieving strong results [9]. Group 3: Contextual Understanding - Llama 4 Scout supports a context window of up to 10 million tokens, setting a new record for open-source models and outperforming competitors like GPT-4o [9]. Group 4: Competitive Landscape - The release of Llama 4 comes amid increasing competition in the open-source model space, particularly from DeepSeek and Alibaba's Tongyi Qianwen series [11][12]. - Meta's previous open-source initiatives, such as Llama 2, have spurred innovation within the developer community, leading to a vibrant ecosystem [11]. - The competitive environment is intensifying, with ongoing advancements in model capabilities and frequent releases from various companies [13].