Mozilla 首份《开源 AI 现状》报告发布：性能差距缩至 3%，但基础设施仍是瓶颈
英文标题：🔶 Mozilla's Inaugural 'State of Open Source AI' Report Is Here
Mozilla 发布首份《开源 AI 现状》报告，基于对全球 950 多名开发者的调查和最新分析指出，开源模型与 ChatGPT、Claude 等顶级闭源系统的性能差距已缩小至 3%，三年内成本最高下降 50 倍。报告称，开源模型约支撑了现实世界 AI 使用量的三分之一，但仅获得 4% 的收入，价值回流至开源生态的机制仍然不足。79% 的开发者正在使用开源模型，但仅 51% 已将其部署到生产环境，低于闭源模型的 63%，凸显部署工具、基础设施和支持体系的缺口。报告还认为，AI Agent 的支撑与编排软件层将成为关键控制点，开源生态需要在基础设施、工具和治理上投入，才能将“开放”转化为可实际使用的能力。
932 B
34.8 KB
43.2 KB
1.3 KB
471 B
1.4 KB
751 B
655 B
638 B
61.6 KB
5.7 KB
175.3 KB
13.0 KB
14.1 KB
1.9 KB
59.3 KB
41.9 KB
中文
开源 AI 已跨越一道门槛。多年来，争论的焦点始终是开放模型能否与闭源模型竞争。这个问题已经不再值得继续争论。今天，Mozilla 发布其首份 《开源 AI 现状》报告。报告基于新的分析以及对全球 950 多名开发者的调查，显示开放模型已不再处于追赶阶段。开源模型与 ChatGPT、Claude 等顶级专有系统之间的性能差距已缩小至仅 3%，而成本在三年间最高下降了 50 倍。
但该报告也揭示了一个更棘手的事实。开放模型如今支撑了约三分之一的现实世界 AI 使用量，却只获得了 4% 的收入。因此，价值虽然已经产生，却没有回流到促成这一切的开源生态系统中。与此同时，地缘政治格局正在快速变化。
亚洲正在加速领先
中国和东亚目前以 89% 的开源 AI 采用率领跑全球，远远领先于西方，并将开源视为国家战略的基石。其他地区的政府也在作出回应：去年新推出了 12 项国家级 AI 战略，目前已有 47 个国家限制关键工作负载在境外进行处理。欧盟、加拿大和印度正通过实际的公共投资支持开放生态系统，将 AI 基础设施视为战略资产。
开发者想要开放模型，但并不总能将其上线。
开发者对开源的需求确实存在，成本和隐私是开发者选择模型时最主要的原因。但希望使用开放模型，与在生产环境中运行它们是两回事。79% 的开发者使用开放模型，但仅 51% 已将其部署到生产环境；相比之下，闭源模型的生产部署比例为 63%。
正如为 Mozilla 开展这项调查的 SlashData 的 Álvaro Ruiz Cubero 所说：“这一差距表明，问题并不单纯在于模型质量，而在于基础设施的缺失。开放模型的部署率几乎不会随着公司规模扩大而提高，这凸显出成熟工具和支持的不足。与此同时，采购方正优先考虑许可条款（31%）和所有权（26%），这表明市场正明显从关注原始能力转向重视控制权和灵活性。”
真正的竞争已转移到模型之外
或许最重要的发现恰恰也是最不容易察觉的：最关键的层并不是模型，而是 Agent 支撑层（agentic harness）——即位于人与模型之间、决定 AI 系统能够看到什么、记住什么以及可以做什么的软件。改变周边软件对性能的影响，可能比更换模型本身更大。谁控制了这一层，谁就控制了 AI 在现实世界中的行为方式。
目前，这一层的构建几乎没有多少防护栏。用户对 AI Agent 请求默认批准的比例高达 93%，而所谓的“同意疲劳”在很大程度上尚未受到关注。
Mozilla 首席技术官 Raffi Krikorian 表示：“开源 AI 已到达一个转折点。重点不再是扩大模型访问权，而是谁拥有塑造、审计和改进它们的权力。如果不投资于开放模型周围的基础设施、工具和治理，我们就可能固化一种只有受限制的闭源 AI 才能规模化发展的体系——而这既不符合公共利益，也不利于在科技政策决策中维护自主权。”
成为所有者，而非租户
开放模型能够提供订阅服务永远无法提供的东西：拥有自己的基础设施，而不是租用他人的基础设施。随着账单不断累积，Microsoft 和 Uber 等公司已经在重新评估自己对付费闭源 AI 工具的依赖。经济模式正在改变，思维方式也在随之转变。
如果你是一名 CTO、政策制定者、开发者或投资者，这份报告值得你阅读。请阅读、分享，并帮助构建相应基础设施，让开放不仅是可能的，更是真正可用的。
English / 原文
Open source AI has crossed a threshold. For years the debate has centered around whether open models could ever compete with closed ones. This is no longer a debate we should be having. Today Mozilla is publishing its inaugural State of Open Source AI report, built on new analysis and a global survey of 950+ developers, showing that open models are no longer playing catch-up. The performance gap with top proprietary systems like ChatGPT and Claude has narrowed to just 3%, while costs have fallen up to 50x in three years.
But the report also surfaces a harder truth. Open models now power roughly a third of real-world AI usage, but capture only 4% of the revenue. So while the value is there, it just isn’t flowing back into the open ecosystem that made it possible. Meanwhile, the geopolitical picture is shifting fast.
Asia is racing ahead
China and East Asia now lead the world in open source AI adoption, at 89%, far ahead of the West, treating open source as a cornerstone of national strategy. Governments elsewhere are responding, with 12 new national AI strategies launched last year, and 47 countries now restrict foreign processing for critical workloads. The EU, Canada, and India are backing open ecosystems with real public investment, treating AI infrastructure as a strategic asset.
Developers want open models, but they can’t always ship them.
For developers, the appetite for open source is there, and cost and privacy rank as developers’ top reasons for choosing models. But wanting to use open models and running them in production are different things. 79% of developers use open models, yet only 51% have deployed them in production, versus 63% for closed models.
As Álvaro Ruiz Cubero of SlashData, which fielded the survey for Mozilla, put it: “this gap indicates that there is not an issue purely of model quality, but of missing infrastructure. Deployment rates for open models barely increase with company size, highlighting a lack of mature tooling and support. At the same time, buyers are prioritising licensing terms (31%) and ownership (26%), signalling a clear shift toward control and flexibility over raw capability.”
The real battle has moved beyond the model
Perhaps the most important finding may be the least obvious. The layer that matters most isn’t the model. It’s the agentic harness: the software between people and models that decides what an AI system can see, remember, and do. Changing the surrounding software can affect performance more than switching the model itself. Whoever controls that layer controls how AI behaves in the world.
Right now, that layer is being built with very few guardrails. Users approve AI agent requests by default up to 93% of the time, with so-called “consent fatigue” going largely unnoticed.
“Open source AI has reached a turning point,” said Raffi Krikorian, Mozilla’s Chief Technology Officer. “It’s no longer about expanding access to models; it’s about who has the power to shape, audit, and improve them. Without investment in the infrastructure, tooling, and governance around open models, we risk locking in a system where only restrictive, closed AI can scale – and that doesn’t serve the public interest, or sovereignty over tech policy decisions.”
Be owners, not renters
Open models offer what a subscription never can: owning your infrastructure instead of renting someone else’s. Companies like Microsoft and Uber are already rethinking their reliance on paid, closed AI tools as the bills add up. The economics are shifting and so is the thinking.
If you’re a CTO, a policymaker, a developer or an investor, this is the report for you. Read it, share it, and help build the infrastructure to make openness usable, not just possible.