支付宝

英媒：美国人工智能公司正受到中国对手的威胁_我的网站

一 |     对开源模型的需求正在飙升          插图：贾里德·巴特曼          2026年7月21日          想想美国顶尖的人工智能开发者们吧。    

    Photo: Global Times
Editor's Note: 

Ahead of the opening of the 2026 World AI Conference and High-Level Meeting on Global AI Governance, Chinese AI start-up Moonshot AI released its Kimi K3 large language model. As the world's largest open-source model by parameter count to date, this launch marks a significant step forward in the development of China's artificial intelligence models.
From the C919 airliner soaring into the skies to Unitree's humanoid robots stealing the show; from DeepSeek pushing the AI frontier to Moonshot AI's landmark unveiling of Kimi K3 - China's wave of homegrown innovations has dominated global headlines in recent years, delivering a steady stream of breakthroughs.
At a meeting in Beijing that brought together the national science and technology award conference, the general assemblies of the members of the Chinese Academy of Sciences (CAS) and the Chinese Academy of Engineering (CAE), and the 11th national congress of the China Association for Science and Technology in July, Chinese President Xi Jinping, also general secretary of the Communist Party of China (CPC) Central Committee and chairman of the Central Military Commission, stressed that the 15th Five-Year Plan period (2026-2030) is a critical phase for tackling tough challenges in building up the country's strength in science and technology.
"We must seize the historic opportunity, rise to the challenges of the times, accelerate efforts to achieve high-level self-reliance and strength in science and technology, and make steady progress toward the 2035 goal of becoming a leading country in science and technology," he said.
In the article "Strive for Greater Strength and Self-Reliance in Science and Technology" included in the fourth volume of Xi Jinping: The Governance of China, Xi pointed out, "Through years of endeavor, our country's overall strength in science and technology has improved substantially. We therefore have a solid foundation, and are fully confident in our ability to seize the opportunities offered by the new revolution in science, technology and industry to achieve greater results." The article also mentioned that "We should participate to the full in global science and technology governance, contribute Chinese wisdom, and shape a philosophy of technology for good purposes, so that science and technology better serve human wellbeing, and enable China's science and technology industry to contribute more to building a global community of shared future." 
In the 27th installment of the special series "Decoding the Book Xi Jinping: The Governance of China," the Global Times, along with the People's Daily Overseas Edition, continues to invite Chinese and foreign scholars, translators of Xi's works, practitioners with firsthand experience, and international readers to focus on the theme of striving for greater strength and self-reliance in science and technology. Together, they share views on China's tech growth, governance principles, and global cooperation in science and technology.
Here are the stories:
1. Readers’ Reflections: China’s innovation success is real global news, landing in the living rooms of Italian retirees who barely use a smartphone: Italian scientist
2. Practitioners' Insights: How did a decade of dedicated effort build China a self-reliant 'digital track'?
3. Translators’ Voices: China's pursuit of strength and self-reliance in science and technology presents an inspiring path
4. Scholars’ Perspectives: National support, optimized evaluation allow young talent to pursue long-term breakthroughs: researcher
    Photo: Global Times
。他们为每一个前沿模型的发布都倾注了无数心血、汗水和巨额资金。然而，还没等他们喘口气，竞争对手就已经迎头赶上。如果只有那些收费的竞争对手紧追不舍，那还算不错。但像Anthropologie和OpenAI这样的公司，还得拼命追赶那些免费提供模型的公司，才能勉强保持领先。

二 |          7月16日，中国初创公司Moonshot AI发布了Kimi K 3模型，其性能仅略逊于Anthropic和OpenAI的最新产品Fable和Sol。该模型的参数（或称“权重”）将于本月晚些时候免费公开，这意味着用户可以下载并在任何支持该功能的服务器上运行它。7月19日，中国互联网巨头阿里巴巴也发布了其Qwen系列模型的新成员预览版，该公司称该模型的性能仅次于Fable。         据研究公司Epoch AI的分析，目前最强大的开源权重模型（通常来自中国）与该技术的前沿水平仅相差数月（见图表）。这些系统的应用正在迅速普及。5月份，OpenRouter（一个热门的模型交易平台）的客户对美国顶级实验室提供的系统的使用量与对中国顶级实验室提供的系统的使用量基本持平，这是根据代币消耗量（业内常用的产出单位）来衡量的。

三 | 次月，美国模型的使用量增长了三分之二，而中国模型的使用量则增长了两倍。         这不仅令美国顶尖的模型开发商感到不安，也令美国政府感到担忧，他们担心美国在人工智能领域的领先地位会因此丧失。

四 | 据说，特朗普政府内部一些人士正在推动采取各种措施来减缓中国模型的传播。然而，美国政府在授予Fable和Sol等先进模型访问权限方面采取的反复无常的做法，也促使人们开始采用中国开发的开源替代模型。         早在K3发布之前，中国的开放权重模型就已引起全球关注。

五 | 中国另一家实验室Z.ai于6月推出的GLM -5.2模型，在诸如软件开发等所谓的智能体任务上，表现优于美国另一家顶级模型开发商谷歌目前提供的任何模型。另外两家中国实验室DeepSeek和MiniMax的模型也广受欢迎。据Firefox浏览器开发商Mozilla的一份报告显示，开放权重模型在前沿推理和复杂智能体任务方面仍处于落后地位，目前已被广泛认为足以胜任许多任务。         它们的成本也低得多——这在企业努力应对人工智能成本不断上涨的问题时是一项重要的优势。

六 | 另一家研究公司 Artificial Analysis 会追踪选定模型在其测试套件中执行每项任务的成本。DeepSeek 提供的最强大的系统平均每次任务的成本为 0.04 美元。尚未开源的K3平均成本为 0.95 美元。Fable 的用户预计需要支付 2.75 美元。         成本差异的部分原因在于，几乎任何云服务提供商都可以出售开放权重模型的访问权限，从而拉低利润率。例如，DeepSeek 提供其 v4 闪存模型的访问权限，价格为每百万代币 0.28 美元，而最便宜的第三方服务提供商的价格为每百万代币 0.18 美元。         此外，许多开源模型在训练时特别注重高效运行，部分原因是由于美国对先进芯片的限制，中国实验室在获取计算能力方面面临挑战。诸如“专家混合模型”之类的创新技术已被开源模型开发者广泛应用，该模型仅激活人工智能系统中与查询相关的部分，DeepSeek 和阿里巴巴等实验室在该技术领域取得了长足进步。

七 |          然而，开源模型的吸引力远不止于此。

八 | 由于美国政府仓促试图阻止除美国公民以外的所有人使用Fable模型，导致该模型在全球范围内停用三周，这促使许多公司重新审视自身对独立实验室的依赖。这促使他们尝试使用开源替代方案，这些方案可以通过多个供应商访问，甚至可以在内部服务器上运行。

九 | Mozilla首席技术官拉菲·克里科里安指出，这一事件犹如一声“警钟”。

Current article:http://ryi.jiubengtaibingpengou.pics/5ditp/20260826/332207.html

Published on:15:42:18