Why open-weight models are crucial for American AI leadership

Artificial Intelligence Cybersecurity Technology Policy & Regulation

人工智能网络安全技术政策与法规

The greatest single potential threat to sustained American artificial intelligence (AI) leadership would be new policies that impede the domestic development of a thriving ecosystem of open-weight AI providers and solutions.

对美国人工智能（AI）持续领导地位的最大单一潜在威胁将是阻碍国内发展繁荣的开放式人工智能提供商和解决方案生态系统的新政策。

To see why, it is helpful to explain what open-weight models are, how they fit within the broader context of AI, and why they are so important for the digital economy.

要了解其中的原因，解释什么是开放权重模型、它们如何适应更广泛的人工智能背景以及为什么它们对数字经济如此重要会有所帮助。

AI systems exist along a spectrum.

人工智能系统存在于一个范围内。

On one end are closed models, such as the various tiers of Anthropic’s Claude and OpenAI’s ChatGPT.

一方面是封闭模型，例如 Anthropic 的 Claude 和 OpenAI 的 ChatGPT 的各个层。

Users can submit queries to these models and get responses, but the detailed inner workings of the models and how they were trained are proprietary.

用户可以向这些模型提交查询并获得响应，但模型的详细内部工作原理以及它们的训练方式是专有的。

At the other end of the spectrum are open-source models , in which information about the training data, the code used to generate weights from the training data, and the resulting weights are publicly disclosed.

另一方面是开源模型，其中有关训练数据的信息、用于从训练数据生成权重的代码以及生成的权重都是公开披露的。

In between these two endpoints are “open-weight” models in which the weights that “ determine how the model interprets input data and generates outputs ” are public, but the methods that were used to generate those weights are not.

在这两个端点之间是“开放权重”模型，其中“确定模型如何解释输入数据并生成输出”的权重是公开的，但用于生成这些权重的方法不是公开的。

An AI ecosystem that lacks robustness in any of these AI categories is deficient, as each type of AI plays a distinct and complementary role.

在任何这些人工智能类别中缺乏稳健性的人工智能生态系统都是有缺陷的，因为每种类型的人工智能都发挥着独特和互补的作用。

Providers of closed models retain full control over how their models are deployed and used, giving them a strong incentive to invest heavily in providing a continuously updated, highly curated set of products that operate within carefully constructed guardrails.

封闭模型的提供商保留对其模型的部署和使用方式的完全控制，这使他们有强烈的动力投入大量资金来提供不断更新、精心策划的产品集，这些产品在精心构建的护栏内运行。

Open-source models are invaluable for researchers and companies seeking a foundation that provides the maximum degree of latitude in developing enhanced, customized in-house solutions.

开源模型对于寻求基础的研究人员和公司来说非常宝贵，该基础可以为开发增强的定制内部解决方案提供最大程度的自由度。

Open-weight models are vital for users who need the flexibility to use or adapt a set of pre-computed model weights but would gain little additional benefit from access to the full set of information about how those weights were derived.

对于需要灵活使用或调整一组预先计算的模型权重的用户来说，开放权重模型至关重要，但从访问有关如何导出这些权重的全套信息中获得的额外好处却很少。

Both open-source and open-weight models are important tools to help the next generation of AI researchers and entrepreneurs learn about AI.

开源和开放权重模型都是帮助下一代人工智能研究人员和企业家了解人工智能的重要工具。

Solutions built on open models can also be much more cost-effective over the longer term than solutions that rely on continual payments to providers of closed models.

从长远来看，基于开放模型构建的解决方案比依赖于向封闭模型提供商持续付款的解决方案更具成本效益。

Discussions about open-weight AI models often raise issues of risks including security.

关于开放权重人工智能模型的讨论经常会引发包括安全在内的风险问题。

But open systems can actually be the foundation for building more secure systems, as they are open to scrutiny by anyone and easier to evaluate.

但开放系统实际上可以成为构建更安全系统的基础，因为它们可以接受任何人的审查并且更容易评估。

This contrasts with closed models, which have inner workings examined only by the people within the companies that developed them.

这与封闭模型形成鲜明对比，封闭模型的内部运作仅由开发它们的公司内部人员检查。

Open-weight models are also criticized because they lack the guardrails that are found in closed models.

开放式重量模型也受到批评，因为它们缺乏封闭式模型中的护栏。

This is a fair critique, but there is a mirroring critique that cuts the other way: The guardrails built into closed models can create their own cybersecurity concerns, because they can make it impossible to use the model to identify and patch security holes in third-party systems.

这是一个公平的批评，但也存在一种反其道而行之的镜像批评：封闭模型中内置的护栏可能会产生自己的网络安全问题，因为它们可能导致无法使用该模型来识别和修补第三方系统中的安全漏洞。

Consider the inadvertent July 2026 intrusion into Hugging Face’s infrastructure by a pair of OpenAI models that escaped their sandbox.

想一想 2026 年 7 月，一对逃离沙箱的 OpenAI 模型无意中入侵了 Hugging Face 的基础设施。

In a post describing how it analyzed the intrusion, Hugging Face explained that the “models we reached for first [to help investigate], Claude Opus and Fable, refused a large part of that work: their safety guardrails treated reverse-engineering an exploit the same as launching one.” Having been unable to make sufficient progress using closed models, Hugging Face then turned to Nvidia’s open weight GLM-5.2 NVFP4 model, which is in turn a “quantized” (i.e., streamlined so that runs faster but at potentially lower precision) version of ZAI GLM-5.2, an open-weight model from Chinese company Z.ai.

在一篇描述如何分析入侵的帖子中，Hugging Face 解释道，“我们首先[帮助调查]的模型 Claude Opus 和 Fable 拒绝了大部分工作：他们的安全护栏将逆向工程漏洞视为与启动漏洞一样。”由于使用封闭模型无法取得足够的进展，Hugging Face 随后转向 Nvidia 的开放权重 GLM-5.2 NVFP4 模型，该模型又是 ZAI GLM-5.2（来自中国公司 Z.ai 的开放权重模型）的“量化”（即简化，以便运行速度更快，但精度可能较低）版本。

The open-weight model provided valuable information in enabling Hugging Face to unravel how the intrusion occurred.

开放权重模型提供了有价值的信息，使 Hugging Face 能够揭示入侵是如何发生的。

Closed AI systems are run on the cloud.

封闭式人工智能系统在云端运行。

This involves a delay as information is sent from an “edge” device to the cloud and back again.

这涉及到信息从“边缘”设备发送到云端并再次返回时的延迟。

For some applications this delay—for example, a few tens of milliseconds—is inconsequential.

对于某些应用程序来说，这种延迟（例如几十毫秒）是无关紧要的。

But for many applications, such as navigating an electric vehicle (EV) or a search and rescue drone, that delay is unacceptable.

但对于许多应用来说，例如电动汽车 (EV) 或搜救无人机的导航，这种延迟是不可接受的。

By contrast, open-weight models (and more specialized systems derived from them) can be run locally on “edge” devices, thereby eliminating the delays inherent in relying on remote computing infrastructure.

相比之下，开放权重模型（以及从中派生的更专业的系统）可以在“边缘”设备上本地运行，从而消除了依赖远程计算基础设施所固有的延迟。

Running models locally also removes the potential operational impediments posed by outages in the cloud or in the communication links between the cloud and edge device.

在本地运行模型还可以消除因云中或云与边缘设备之间的通信链路中断而造成的潜在操作障碍。

While much of the discussion about AI focuses on models, that isn’t the full story.

虽然关于人工智能的大部分讨论都集中在模型上，但这并不是故事的全部。

How an AI model operates in practice is also determined by the particular scaffolding and harness that surrounds it.

人工智能模型在实践中的运作方式也取决于其周围的特定脚手架和工具。

An example is Claude Code.

克劳德代码就是一个例子。

Claude is an underlying family of models that are useful for an enormous spectrum of tasks.

Claude 是一个基础模型系列，可用于多种任务。

Claude Code adds a harness to this framework to create a product that is optimized for producing computer code.

Claude Code 向该框架添加了一个工具，以创建针对生成计算机代码进行优化的产品。

While Claude Code is extremely good, it is a harness built on a closed system.

虽然 Claude Code 非常好，但它是建立在封闭系统上的工具。

Providers of closed systems will be limited in how many harnesses they will build for their models.

封闭系统的提供商将受到为其模型构建的线束数量的限制。

By contrast, open-weight systems allow anyone to develop (and potentially monetize) a harness to optimize the model for use in a specific task.

相比之下，开放权重系统允许任何人开发（并可能货币化）工具来优化特定任务中使用的模型。

More harnesses mean that AI can be useful to more people and for a greater variety of applications.

更多的利用意味着人工智能可以为更多的人和更多种类的应用带来帮助。

The best American response to robust overseas competition is to innovate faster.

美国对激烈的海外竞争的最佳反应是加快创新。

Bans lower the incentives to innovate and invite retaliation from other nations.

禁令降低了创新的动力，并招致其他国家的报复。

The market for EVs, where punitive U.S. tariffs on Chinese EVs amount to a de facto ban, furnishes a useful analogy.

电动汽车市场提供了一个有用的类比，美国对中国电动汽车征收惩罚性关税相当于事实上的禁令。

Protectionist U.S. EV policies have resulted in higher prices and less choice for American EV buyers, less EV innovation in the U.S., and the rapid acceleration of market dominance by Chinese EVs in countries outside the U.S.

美国电动汽车保护主义政策导致价格上涨、美国电动汽车买家的选择减少、美国电动汽车创新减少以及中国电动汽车在美国以外国家的市场主导地位迅速加速。

While calls in U.S. policy circles to place restrictions on Chinese open-weight AI models typically cite (with little to no supporting evidence) alleged threats to national security, that sidesteps the fact that Chinese open-weight models are extraordinarily good.

尽管美国政策圈呼吁对中国开放权重人工智能模型施加限制，但通常会引用（几乎没有任何支持证据）所谓的对国家安全的威胁，但这回避了中国开放权重模型非常好的事实。

Unless artificial trade barriers intervene, people around the world will seek out and adopt the best models, regardless of their country of origin.

除非人为的贸易壁垒干预，否则世界各地的人们都会寻求并采用最好的模式，无论他们来自哪个国家。

If U.S. companies can build open-weight models that are on par with or better than the best Chinese models, the U.S. models will experience wide global adoption.

如果美国公司能够构建与中国最佳模型相当或更好的开放重量模型，那么美国模型将在全球范围内得到广泛采用。

Startups in the open-weight space include Poolside, Reflection, and Hugging Face in the U.S.; DeepSeek, Moonshot AI, Z.ai (also called Zhipu AI) in China; and Mistral in France.

开放重量级领域的初创公司包括美国的 Poolside、Reflection 和 Hugging Face；中国的DeepSeek、Moonshot AI、Z.ai（也称为Zhipu AI）；和法国的米斯特拉尔。

These and other companies in the field are using a variety of approaches to generate revenue, with offerings including customer-specific model adaptation, hosting services, and agents that “ plan and execute multi-step tasks ” within an enterprise.

这些公司和该领域的其他公司正在使用各种方法来创收，提供的产品包括客户特定的模型适应、托管服务以及在企业内“规划和执行多步骤任务”的代理。

To date, venture capital investments and valuations in the open-weight space are far smaller than with the leading closed model providers, OpenAI and Anthropic.

迄今为止，开放权重领域的风险投资和估值远远小于领先的封闭模型提供商 OpenAI 和 Anthropic。

But AI changes fast, and while the analogy is imperfect, Linux and Android provide important precedents showing that open systems can be the foundation for highly profitable products and services.

但人工智能变化很快，虽然类比并不完美，但 Linux 和 Android 提供了重要的先例，表明开放系统可以成为高利润产品和服务的基础。

If American AI is dominated by a small handful of closed model providers, the result will be an overconcentration of power.

如果美国人工智能由少数封闭模型提供商主导，结果将是权力过度集中。

That raises many concerns, including the risk of regulatory capture and a set of resulting policies that impede new market entrants.

这引发了许多担忧，包括监管捕获的风险以及由此产生的一系列阻碍新市场进入者的政策。

A far better approach is to foster an environment in which closed model providers and open model providers can both thrive, offering complementary products and services.

更好的方法是营造一个封闭模式提供商和开放模式提供商都能蓬勃发展的环境，提供互补的产品和服务。

Computing in the age of AI is creating extraordinary opportunities for innovation across fields including health care delivery, education, pharmaceuticals, manufacturing, and many more.

人工智能时代的计算正在为医疗保健、教育、制药、制造等领域的创新创造非凡的机会。

The best way to allow 21st century American ingenuity, amplified by AI, to reach its full potential is to promote growth of a diversified ecosystem with global leadership across all categories of AI.

让 21 世纪美国的创造力在人工智能的推动下充分发挥潜力，最好的方法是促进多元化生态系统的发展，并在所有人工智能领域保持全球领先地位。