AI 人才争夺战愈演愈烈，就算是顶级大厂，如果没有"护城河"，也留不住人。 据外媒 Business Insider 最新消息，曾在开源大模型圈子里一度领跑的 Meta，如今正面临严重的人才流失。在 Llama 模型最初的 14 位核心作者 中，已有 11 位离职。有的自立门户，有的跳槽去了竞争对手。 这波"出走潮"也让外界再次把目光投向 Meta。毕竟他们曾豪赌元宇宙，四年"烧掉"450 亿美元，却被直指至今几乎未见显著成效。现在 AI 项目 也出问题了，不少人开始质疑：Meta 还行不行？为什么留不住顶尖 AI 人才？它的创新能力，还能支撑它在这场 AI 竞赛中跑多远？ Llama 论文的 14 位作者，已有 11 人离开 Meta 回头看 2023 年那篇引发轰动的 Llama 论文，共署名 14 位研究者。短短两年，Meta 只留下了其中三位：研究科学家 Hugo Touvron、研究工程师 Xavier Martinet 和项目负责人 Faisal Azhar。 论文地址：https://arxiv.org/pdf/2302.13971 其他 11 人，大多已经离开，分散到了全球多家科技公司，有的还 ...

Investment Rating - The report maintains a "Recommendation" rating for the media industry, expecting the industry index to rise more than 5% over the next 3-6 months compared to the benchmark index [49]. Core Insights - The media sector is experiencing a resurgence in the film and gaming markets, driven by advancements in AI applications and cultural confidence stemming from popular IPs like "Nezha" [3][6]. - The report highlights the importance of AI applications in reshaping the industry landscape, with a focus on public cloud value reconstruction and the return to growth for related companies [6][9]. - The gaming market shows strong performance, with Tencent and NetEase leading in iOS sales rankings, indicating a healthy competitive environment [16][19]. - The film market is recovering, with total box office revenue reaching 239 billion yuan as of May 9, 2025, which is approximately 102% of the 2019 level [20][23]. Market Performance Review - The media (Shenwan) index rose by 1.40% last week, underperforming the CSI 300 index, which increased by 2.00%, resulting in a relative underperformance of 0.61% [9][10]. - The report notes that the media sector ranks 24th among all sectors in terms of performance [9]. Industry Highlights - The report emphasizes the potential of AI applications in various sectors, including gaming and education, suggesting a focus on companies like Huya, Giant Network, and Perfect World for gaming, and New Oriental and TAL Education for education [6][30]. - The film industry is expected to benefit from a strong pipeline of upcoming releases, with several key films set to debut in mid-May [29][30]. Company Announcements - Perfect World announced a 2025 employee stock ownership plan, indicating a commitment to employee engagement and retention [36]. - Wanda Film disclosed plans for a share reduction by a major shareholder, which may impact market sentiment [38].

Core Viewpoint - The domestic large model industry is attempting to break through its current challenges and reconstruct a new order, but the unstable market environment poses significant risks [1] Group 1: Open Source Trends - DeepSeek has disrupted the industry's perception of open-source models, prompting OpenAI's CEO to reconsider the validity of open-source strategies [1] - Domestic large model companies like Alibaba, Baidu, and SenseTime are accelerating their open-source initiatives [1] - Open-source is seen as a key strategy to reduce dependency on foreign software and hardware, but the commercial viability of open-source projects remains complex [2][5] Group 2: Challenges in Implementation - Developers face significant technical adaptation and maintenance costs, despite open-source models lowering the technical barrier [4] - The integration of large models into existing systems requires extensive customization, which can be resource-intensive for companies [4] - The complexity of data acquisition, cleaning, and labeling poses additional challenges for businesses, particularly small and medium-sized enterprises [4] Group 3: Investor Sentiment - Investors are cautious about the open-source model due to the unclear profitability and traditional software sales evaluation methods not being applicable [5] - The potential for significant financial loss if investments in proprietary models are undermined by open-source alternatives is a concern for investors [4][5] Group 4: Business Models - Chinese large model companies are adopting a "free-to-use plus value-added services" model to build a commercial framework around open-source models [6][8] - Companies like Baidu are leveraging their cloud services to monetize the usage of their open-source models, creating a win-win situation for developers and the company [8] - The success of open-source models may depend more on the quality of cloud services than on the models themselves, as seen in the strategies of Meta and Hugging Face [9][10] Group 5: Future Outlook - Open-source is viewed as a pathway for the Chinese large model industry to overcome technological barriers, but commercial sustainability is equally important [10] - The increasing tariff barriers from the U.S. add pressure to the large model industry, making the choice of cloud platforms more critical than the open-source models themselves [10]

Group 1 - The revised Marriage Registration Regulations will take effect on May 10, 2025, eliminating the requirement to present a household registration book for marriage and divorce [3] - The new regulations include three main aspects: expanding marriage and family service content, implementing nationwide marriage registration, and optimizing marriage registration services [3] - The marriage registration authority is prohibited from charging fees for processing marriage and divorce registrations [3] Group 2 - Xiamen Jihong Technology Co., Ltd. has passed the listing hearing on the Hong Kong Stock Exchange, with China International Capital Corporation and China Merchants Jinling International serving as joint sponsors [2] - Panasonic Group announced plans to lay off 10,000 employees globally, with 5,000 from Japan and 5,000 from overseas, during the fiscal year 2025-2026 [6] - Ningde Times is reportedly seeking to raise at least $4 billion through a Hong Kong listing [6] Group 3 - The Italian company Moltiply has filed a lawsuit against Google's parent company Alphabet, seeking €2.97 billion ($3.34 billion) in damages for abusing its market dominance [5] - The U.S. tariff war has significantly increased costs for American companies, with one bicycle manufacturer reporting a nearly threefold increase in wheel costs due to tariffs [6] - The recent divorce of the controlling shareholder of Zhu Cheng Technology involves the transfer of approximately 3.81 million shares, valued at around 3.81 million yuan [6]

Investment Rating - The report indicates that Alibaba's Qwen currently ranks at the top of the open-source model rankings, with expectations for continued leadership in model capability and ecosystem monetization [2]. Core Insights - The report highlights a surge in domestic open-source models, with significant releases from Alibaba, Xiaomi, and DeepSeek, showcasing advancements in large language models (LLMs) [1][8]. - Alibaba's Qwen-3 series demonstrates substantial performance improvements, achieving 10-30% accuracy gains on various benchmarks and enhancing inference speed by 20-40% [9][12]. - Xiaomi's MiMo model, with 7 billion parameters, excels in reasoning and code generation tasks, outperforming larger proprietary models through innovative training strategies [10][12]. - DeepSeek's Prover-V2-671B model shows strong performance in formal logic reasoning, indicating a strategic focus on specialized AI applications [11][12]. - The report anticipates that as more domestic models are released, the industry may face challenges related to homogenization and competition, pushing for more customized solutions in vertical industries [5]. Summary by Sections Alibaba Qwen-3 - The Qwen-3 series includes models ranging from 1.5 billion to 72 billion parameters, designed for various inference needs, with notable performance enhancements over previous generations [9]. - Deployment costs are significantly lower, requiring only 4 H20 GPUs for full-capacity operation, which is advantageous compared to similar models from OpenAI and Grok [2][12]. Xiaomi MiMo - MiMo's training involved 25 trillion tokens and innovative mechanisms to improve training efficiency, achieving a 2.29x increase in training speed and a 1.96x acceleration in verification processes [10]. DeepSeek-Prover-V2-671B - This model excels in mathematical theorem proving, particularly in formal logic, and serves as a precursor to DeepSeek's upcoming models, reflecting the company's commitment to advancing AI capabilities [11]. Industry Trends - The report suggests that the next phase for open-source models will involve customization based on user data and feedback, aiming to establish long-term barriers and user loyalty in specific industries [5].

鱼羊 一水 发自 凹非寺 量子位 | 公众号 QbitAI 开源大模型新王者，正在受到空前关注。 Qwen3预告一出，直接开启不眠夜模式。 △ 来自编辑部本部 等到深夜正式上线并宣布登顶全球最强开源模型，更是瞬间引爆全网热议。 | | | Hope you enjoy our new models! | | | | | | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | 22B | Qwen3-32B Dense | OpenAl-o1 2024-12-17 | Deepseek-R1 | Grok 3 Beta BB Think | QwQ-32B | Qwen3-4B Dense | Qwen2.5-72B-Instruct | Gemma3-27BIT | | | 93.8 | 92.1 | 93.2 | | 89.5 | 76.6 | 81.2 | 86.8 | | | 81.4 | 74.3 | 79.8 | 83.9 | 79.5 | 73.8 | 18.9 | 32.6 | | | 72.9 | 79.2 | 70.0 | ...

阿里新一代的大模型 Qwen 3 今早发布，新旗舰 Qwen3-235B-A22B 的评测成绩，和 DeepSeek R1、Grok-3、Gemini-2.5-Pro 不相上下。这一代全系列模 型都支持混合推理，对 Agent 的支持也上了新台阶。 随着 Qwen 2.5 和 3 的发布，全球的开源模型生态也呈现了一种新形态：以 DeepSeek+Qwen 的中国开源组合，取代了过去 Llama 为主，Mistral 为辅的开 源生态。Qwen 系列的衍生模型目前已经是 HuggingFace 上最受欢迎的开源模型，衍生模型的数量也超过了 Llama 系列。而 DeepSeek 对于开源模型生态 的冲击和贡献，也有目共睹。 与大模型六小龙相比，主打开源的 Qwen 和 DeepSeek 无疑在国际市场赢得了更多开发者和创业者的关注，来自开源社区的代码贡献、更多优秀微调版本 的出现，也在以另外一种方式推动模型能力的进步。 可以说， 开源，正在成为中国大模型公司进入全球市场的最佳路径。 而对阿里云来说，Qwen+阿里云的配合，「模型-云-行业应用」的打法，走出了国内 MaaS 模式的新方向，也在很大程度上降低了国 ...

Core Insights - The China Academy of Information and Communications Technology (CAICT) has released a report highlighting exemplary cases of "Open Source Large Models+" software innovation applications, with Zhiyuan Interconnect recognized as a benchmark in this field [1][3] - The report focuses on the practical implementation of artificial intelligence technologies, selecting benchmarks based on technological breakthroughs, scene innovation, and ecological synergy [3] Group 1: Company Innovations - Zhiyuan Interconnect has developed the AI-COP intelligent collaborative operation platform, integrating "large models + vertical domain models + scene intelligent agents" to create a replicable and scalable industry paradigm [3][4] - The company has launched the "CoMi Family" of intelligent agent products, which combines mainstream AI large models with self-developed vertical domain models, enhancing capabilities from single-process tools to multi-task AI agents [4] Group 2: Product Offerings - The CoMi Family features over ten vertical scene intelligent agents tailored for diverse business scenarios, such as enterprise intelligent inquiry, collaborative work assistants, and contract risk assistants, aimed at improving organizational efficiency and decision-making quality [5] - The intelligent agents utilize data interaction analysis to provide instant and accurate query results, significantly accelerating decision-making processes and breaking the constraints of traditional data querying methods [5][6] Group 3: Market Applications - The company has introduced a one-stop enterprise AI service platform, Zhihuiquan, which integrates over 50 mainstream large models, supporting private deployment and adaptation for various industries, including finance and manufacturing [5] - The applications cover multiple scenarios such as AI comprehensive portals, intelligent documents, and smart data analysis, contributing to quality improvement, cost reduction, and risk mitigation for enterprises [6]

Group 1 - Beijing is actively building a "global open-source capital," focusing on open-source and innovation, aiming to create an open-source innovation ecosystem and optimize infrastructure [1] - Open-source technology has penetrated various industries, including automotive and robotics, with companies like Li Auto and the Beijing Humanoid Robot Innovation Center leading initiatives [1] - The establishment of cross-industry open-source collaborative organizations is recommended to standardize common technologies and create a modular "technology Lego" [1][3] Group 2 - Beijing Zhiyuan Huazhang Technology Co., Ltd. has released a series of open-source models, including 32B and 9B parameter models, and aims to promote AI accessibility [2] - The Beijing Artificial Intelligence Industry Investment Fund has announced an additional investment of 200 million yuan in Zhiyuan to support open-source model development [2] - Zhiyuan's Z Fund will invest 300 million yuan to support global AI open-source community development, allowing startups to apply for funding based on open-source models [2] Group 3 - The core value of open-source communities lies in reducing redundant work and resource waste, fostering collaboration among developers to create superior systems [3] - Open-source large models demonstrate significant advantages in gathering global developer resources, optimizing model capabilities, and adapting to diverse scenarios [3] - Companies can achieve "coexistence in competition" through open-source by sharing non-core technologies, collaborating on innovation projects, and contributing to open-source community development [3] Group 4 - The strategy to deepen open-source initiatives includes organizing the release of RISC-V processor cores and various models, aiming to establish a strong open-source foundation [4]

一提到"大模型"，很多人的第一反应往往是那个既能聊天，又会写代码、画画的"模型本身"。但其 实，大模型远不止是一个"能输出结果的程序"这么简单，其背后有一整套复杂而庞大的技术体系作为 支撑：从大规模、高质量、多样化的数据，到先进的模型架构与训练策略，再到推理部署、资源调度 等支撑落地的系统能力，以及不可或缺的科学评测机制。大模型更像是一个由模型、数据、系统、评 测平台 等多要素构成的"技术共同体"，而非单一模块的堆叠。 如今在闭源技术壁垒与高昂商用门槛的对比下，开源大模型正迅速崛起，成为推动 AI 技术普惠化的 重要力量。但面对层出不穷的开源 AI 模型技术，我们该如何选型？不同的模型技术体系又各有怎样 的优势与短板？ 在这一背景下，为系统呈现全球大模型生态的开源发展现状，CSDN 联合多家机构于 4 月 18 日在 2025 全球机器学习技术大会（ML-Summit 2025）现场重磅发布《AI 大模型技术体系综合开源影响 力榜单》，全面评估全球范围内开源大模型技术体系的贡献与影响力，旨在为行业提供参考坐标，推 动开源创新持续前行。 注：这里大模型是指 主要包括 decoder-only 以来的模型结构，包 ...