The Path to an Open AI Future: Why Llama 3.1 Represents a Turning Point

Introduction: A Historical Parallel

　　In the formative era of high-performance computing, the industry's trajectory was dominated by proprietary, closed-source Unix systems developed by major technology firms. The prevailing wisdom held that such complex, advanced software could only be built within the confines of a tightly controlled, corporate environment. The notion of a collaborative, open alternative seemed implausible. However, the rise of open-source Linux fundamentally altered this landscape. Initially gaining traction due to its affordability and the freedom it offered developers to modify the source code, Linux evolved over time to surpass its proprietary counterparts in sophistication, security, and the breadth of its supporting ecosystem. Today, it is the foundational bedrock of cloud infrastructure and powers the majority of mobile devices, delivering superior products and innovation that benefit everyone.

在高性能计算的早期，行业的发展轨迹由各大科技公司开发的专有、闭源 Unix 系统所主导。当时的普遍观点认为，如此复杂、先进的软件只能在严格控制的企业环境中构建。协作式、开放式的替代方案似乎难以想象。然而，开源 Linux 的崛起从根本上改变了这一格局。最初，Linux 因其经济性和为开发者提供的修改源代码的自由而获得关注，随后逐渐发展，在复杂性、安全性以及支持生态系统的广度上超越了其专有对手。如今，它是云基础设施的基石，并为大多数移动设备提供动力，为所有人带来了更优质的产品和创新。

　　I believe artificial intelligence is poised to follow a remarkably similar evolutionary path. Currently, several leading technology companies are developing state-of-the-art closed models. Yet, the open-source community is rapidly closing the performance gap. Just last year, models like Llama 2 were competitive with a prior generation of frontier models. This year, Llama 3 has demonstrated competitiveness with the most advanced systems and leads in specific benchmarks. Looking ahead, we anticipate that future iterations of Llama will set the industry standard for capability. Even today, the Llama family already establishes leadership in the critical dimensions of openness, modifiability, and cost efficiency.

我相信人工智能正准备好遵循一条极其相似的发展道路。目前，几家领先的科技公司正在开发最先进的闭源模型。然而，开源社区正在迅速缩小性能差距。就在去年，像 Llama 2 这样的模型只能与前沿模型的上一代产品竞争。今年，Llama 3 已经展现出与最先进系统竞争的实力，并在特定基准测试中领先。展望未来，我们预计 Llama 的未来迭代版本将为行业的能力标准树立标杆。即使在今天，Llama 系列已经在开放性、可修改性和成本效益这些关键维度上确立了领导地位。

Announcing Llama 3.1: A New Frontier for Open Source AI

　　Today, we are taking a significant stride toward establishing open-source AI as the industry standard. We are releasing Llama 3.1 405B, the first open-source AI model that operates at the true frontier of capability. Alongside it, we are introducing new and improved versions of Llama 3.1 70B and 8B models. Beyond offering significantly superior cost-to-performance ratios compared to closed models, the open nature of the 405B model will make it the premier choice for fine-tuning and for the process of distilling knowledge into smaller, more efficient models.

今天，我们朝着确立开源 AI 为行业标准的目标迈出了重要一步。我们正式发布 Llama 3.1 405B，这是首个在真实能力前沿运行的开源 AI 模型。与此同时，我们推出了新版且性能提升的 Llama 3.1 70B 和 8B 模型。除了相比闭源模型提供显著更优的性价比之外，405B 模型的开源特性将使其成为微调以及将知识提炼到更小、更高效模型这一过程的首选。

Building a Robust Ecosystem

　　Our commitment extends beyond model releases. We are collaborating with a diverse array of industry partners to cultivate a comprehensive and thriving ecosystem.

• Cloud & Platform Providers: Amazon, Databricks, and NVIDIA are launching full suites of services to support developers in fine-tuning and distilling their own custom models.

• Inference Specialists: Innovators like Groq have built low-latency, cost-effective inference serving infrastructure optimized for all the new Llama 3.1 models.

• Broad Availability: The models will be accessible on all major cloud platforms, including AWS, Azure, Google Cloud, Oracle Cloud, and more.

• Enterprise Enablement: Companies like Scale AI, Dell, Deloitte, and others stand ready to assist enterprises in adopting Llama and training bespoke models with their proprietary data.

　　As this community expands and more companies develop novel services atop this foundation, we can collectively solidify Llama as the industry standard, thereby democratizing the benefits of advanced AI for all.

我们的承诺不止于模型发布。我们正与众多行业伙伴合作，共同培育一个全面而繁荣的生态系统。

• 云与平台提供商：亚马逊、Databricks 和英伟达正在推出全套服务，以支持开发者微调和提炼他们自己的定制模型。

• 推理专家：像 Groq 这样的创新者已经构建了针对所有新 Llama 3.1 模型优化的低延迟、高性价比的推理服务基础设施。

• 广泛可用性：这些模型将在所有主要云平台上可用，包括 AWS、Azure、Google Cloud、Oracle Cloud 等。

• 企业赋能：Scale AI、戴尔、德勤等公司已准备就绪，可协助企业采用 Llama 并使用其专有数据训练定制模型。

随着这个社区的壮大以及更多公司在此基础之上开发新服务，我们可以共同将 Llama 确立为行业标准，从而让先进 AI 的益处惠及所有人。

　　Meta is deeply committed to the open-source AI paradigm. In the following sections, I will outline why open source constitutes the optimal development stack for developers, why open-sourcing Llama aligns with Meta's strategic interests, and why an open-source approach is fundamentally beneficial for global progress and is therefore a sustainable platform for the long term.

Meta 坚定致力于开源 AI 范式。在接下来的部分，我将阐述为何开源是开发者的最佳开发堆栈，为何开源 Llama 符合 Meta 的战略利益，以及为何开源方法从根本上有利于全球进步，并因此是一个可持续的长期平台。

Why Open Source AI Is Good for Developers

　　In my discussions with developers, CEOs, and government officials worldwide, several consistent themes emerge regarding their needs from AI technology:

在我与全球开发者、CEO 和政府官员的讨论中，关于他们对 AI 技术的需求，出现了几个一致的主题：

　　1. The Need for Customization and Control Every organization possesses unique requirements, best addressed by models of varying sizes that are trained or fine-tuned on specific, proprietary datasets. Simple on-device tasks or classification may necessitate small models, while complex reasoning demands larger ones. With open models like Llama 3.1, you can take the most advanced foundation, continue pre-training or fine-tuning it with your own data, and subsequently distill it down to your optimal size—all while maintaining complete data privacy, as neither Meta nor any third party needs to see your sensitive information.

1. 定制与控制的需求 每个组织都有独特的需求，最好通过在不同规模、使用特定专有数据集进行训练或微调的模型来解决。简单的设备端任务或分类可能需要小模型，而复杂的推理则需要大模型。借助像 Llama 3.1 这样的开源模型，你可以基于最先进的基础模型，使用自己的数据进行持续的预训练或微调，随后将其提炼至最优规模——同时保持完全的数据隐私，因为 Meta 或任何第三方都无需接触你的敏感信息。

　　2. Avoiding Vendor Lock-in Many organizations are wary of becoming dependent on models they cannot independently run, audit, or control. They seek to avoid scenarios where a closed-model provider can unilaterally alter the model's behavior, change terms of service, or even discontinue service. Furthermore, they wish to avoid being locked into a single cloud provider that holds exclusive rights to a model. Open source fosters a broad ecosystem of companies offering compatible toolchains and services, ensuring portability and freedom of choice.

2. 避免供应商锁定 许多组织警惕依赖于他们无法独立运行、审计或控制的模型。他们希望避免闭源模型提供商单方面改变模型行为、更改服务条款甚至停止服务的场景。此外，他们也希望避免被绑定在某个对模型拥有独家权利的单一云提供商上。开源培育了一个由提供兼容工具链和服务的公司组成的广泛生态系统，确保了可移植性和选择自由。

　　3. Ensuring Data Security and Privacy Organizations handling sensitive data—whether in healthcare, finance, or government—cannot risk transmitting it to external cloud APIs for processing by closed models. Others simply lack trust in the data stewardship practices of closed-source providers. Open source directly addresses these concerns by enabling models to be deployed and run within an organization's own secure infrastructure, be it on-premises or in a private cloud. It is also widely acknowledged that open-source software tends to achieve higher security standards due to its transparent development process and extensive peer review.

3. 确保数据安全与隐私 处理敏感数据的组织——无论是在医疗、金融还是政府领域——不能冒险将其传输到外部云 API 由闭源模型处理。其他组织则根本不信任闭源提供商的数据管理实践。开源通过使模型能够在组织自己的安全基础设施（无论是本地还是私有云）内部署和运行，直接解决了这些担忧。人们也普遍认为，由于其透明的开发过程和广泛的同行评审，开源软件往往能达到更高的安全标准。

　　4. Demanding Efficiency and Affordability Cost-effectiveness is paramount for scaling AI applications. Developers can run inference with Llama 3.1 405B on their own infrastructure at approximately 50% of the cost of using comparable closed models like GPT-4o, for both user-facing applications and offline batch processing tasks.

4. 要求效率与可负担性 成本效益对于扩展 AI 应用至关重要。开发者可以在自己的基础设施上运行 Llama 3.1 405B 进行推理，其成本大约仅为使用 GPT-4o 等类似闭源模型的 50%，无论是面向用户的应用还是离线批处理任务。

　　5. Investing in a Long-Term Standard Many recognize that the pace of innovation in open-source AI is accelerating faster than in the closed model domain. They seek to build their critical systems on an architectural foundation that promises not just competitiveness today, but sustained advantage and community-driven evolution for the future.

5. 投资于长期标准 许多人认识到，开源 AI 的创新速度正在超过闭源模型领域。他们希望在这样一个架构基础上构建其关键系统：它不仅承诺当下的竞争力，更承诺未来的持续优势以及社区驱动的演进。

Why Open Source AI Is Good for Meta

　　Meta's core business model revolves around creating the best possible experiences and services for people. To achieve this, we must guarantee perpetual access to the best available technology and avoid entrapment within a competitor's closed ecosystem, where our ability to innovate could be arbitrarily restricted.

Meta 的核心商业模式围绕着为用户创造最佳的体验和服务。为了实现这一目标，我们必须保证永远能够获取最先进的技术，并避免陷入竞争对手的封闭生态系统中，因为在那里我们的创新能力可能会受到任意限制。

　　A formative experience for me has been navigating the constraints imposed by building services on platforms like Apple's iOS. The combination of fees, arbitrary rules, and the blocking of product innovations makes it evident that Meta—and countless other companies—could deliver vastly superior services if we were not hindered by the strategic limitations enforced by platform gatekeepers. Philosophically, this is a primary reason for my strong conviction in building open ecosystems for the next generation of computing, encompassing both AI and augmented/virtual reality (AR/VR).

对我而言，一个塑造性的经历是在苹果 iOS 等平台上构建服务时所面临的限制。费用、任意规则以及对产品创新的阻碍结合在一起，清楚地表明，如果 Meta 以及无数其他公司不受平台守门人强加的战略限制的阻碍，我们本可以提供卓越得多的服务。从理念上讲，这是我坚信为下一代计算（包括 AI 和增强/虚拟现实 AR/VR）构建开放生态系统的一个主要原因。

　　A common question is whether I am concerned about relinquishing a technical advantage by open-sourcing Llama. This perspective, I believe, overlooks the broader strategic picture for several reasons:

一个常见的问题是，我是否担心通过开源 Llama 会丧失技术优势。我认为，这种观点忽略了更广泛的战略图景，原因如下：

　　First, to ensure our long-term access to leading technology and avoid lock-in, Llama must evolve into a complete ecosystem encompassing tools, efficiency enhancements, silicon optimizations, and integrations. If Meta were the sole entity utilizing Llama, this rich ecosystem would never materialize, leaving us in a position no better than the proprietary Unix variants of the past.

首先，为了确保我们长期获得领先技术并避免被锁定，Llama 必须发展成为一个完整的生态系统，包含工具、效率提升、芯片优化和集成。如果 Meta 是唯一使用 Llama 的实体，这个丰富的生态系统将永远无法形成，使我们陷入的境地不会比过去的专有 Unix 变体更好。

　　Second, I anticipate that AI development will remain intensely competitive. Open-sourcing a given model does not equate to surrendering a massive, enduring advantage, as the next generation of models is always on the horizon. The path for Llama to become the industry standard is through consistent competitiveness, superior efficiency, and unwavering openness across successive generations.

其次，我预计 AI 发展将保持激烈竞争。开源一个特定模型并不等同于放弃一个巨大而持久的优势，因为下一代模型总是在酝酿之中。Llama 成为行业标准的途径是通过持续的竞争力、卓越的效率以及历代产品坚定不移的开放性。

　　Third, a fundamental distinction between Meta and closed-model providers is that selling access to AI models is not our business model. Therefore, releasing Llama openly does not undermine our core revenue streams, sustainability, or capacity to invest in research—unlike the predicament it creates for companies whose primary product is a proprietary AI API. (This economic reality is a key driver behind lobbying efforts by some closed providers against open-source AI.)

第三，Meta 与闭源模型提供商之间的一个根本区别在于，出售 AI 模型的访问权限并非我们的商业模式。因此，公开释放 Llama 并不会损害我们的核心收入流、可持续性或研究投资能力——这与对那些主要产品是专有 AI API 的公司所造成的困境不同。（这一经济现实是一些闭源提供商游说反对开源 AI 的关键驱动力之一。）

　　Finally, Meta has a proven, successful history with open source. Initiatives like the Open Compute Project (OCP), where we opened our server, network, and data center designs, have saved the company billions of dollars as supply chains standardized around these specifications. We have also reaped immense benefits from the community's innovations by open-sourcing foundational tools like PyTorch and React. This long-term, ecosystem-oriented strategy has consistently yielded positive outcomes for Meta.

最后，Meta 在开源方面有着经过验证的成功历史。像开放计算项目这样的举措，我们开源了服务器、网络和数据中心设计，随着供应链围绕这些规范实现标准化，为公司节省了数十亿美元。通过开源 PyTorch 和 React 等基础工具，我们也从社区的创新中获得了巨大收益。这种长期的、面向生态系统的战略一直为 Meta 带来积极的成果。

Why Open Source AI Is Good for the World

　　I am convinced that open source is indispensable for shaping a positive AI future. AI holds more potential than any contemporary technology to enhance human productivity, unleash creativity, improve quality of life, accelerate economic growth, and unlock breakthroughs in medicine and scientific research. An open-source approach is critical to ensuring that these benefits are broadly distributed.

我深信，开源对于塑造一个积极的 AI 未来是不可或缺的。AI 比任何当代技术都更具潜力，能够提升人类生产力、释放创造力、改善生活质量、加速经济增长，并在医学和科学研究中实现突破。开源方法对于确保这些益处得到广泛分配至关重要。

　　Open source will democratize access to AI's opportunities, preventing the concentration of immense power in the hands of a few corporations. It will foster global innovation, as researchers and developers everywhere can build upon, audit, and improve the underlying technology. This transparency and collaborative effort are essential for building trust, ensuring safety, and aligning AI's development with the broad interests of humanity. By choosing an open path, we are investing in a platform for progress that is inclusive, resilient, and built to last.

开源将使 AI 机会民主化，防止巨大的权力集中在少数公司手中。它将促进全球创新，因为世界各地的研究人员和开发者都可以在此基础上进行构建、审计和改进。这种透明度和协作对于建立信任、确保安全以及使 AI 的发展与人类的广泛利益保持一致至关重要。通过选择开放的道路，我们正在投资于一个包容、有韧性且为持久而建的进步平台。