Mark Zuckerberg says Meta will not lean on open-source AI models for now, citing “risk”

Meta CEO Mark Zuckerberg said Meta cannot yet rely on open-source AI models to meet its ambitions, arguing that the current generation of publicly available frontier models comes with material risks. The remarks, reported by Yahoo Finance, frame open-source as a tool that can be useful, but not one that Meta is prepared to treat as a dependable foundation for its most advanced work right now.

Zuckerberg’s position was also tied to Meta’s broader approach to artificial intelligence development. He indicated that Meta intends to keep building its own frontier models instead of outsourcing the core direction to third-party open-source alternatives, according to the reported comments.

The reported quotes use caution rather than a blanket dismissal. Zuckerberg characterized the situation as one where there is some “risk,” implying that model reliability, performance consistency, or other deployment considerations may be harder to manage with open-source systems today than with models built and controlled in-house.

Meta did not lay out specific categories of risk in the Yahoo Finance report. It did not provide concrete examples, such as failures in safety controls, limitations in evaluation results, or operational issues during deployment. As a result, the exact concern remains unspecified in the available account.

Meta’s stance matters because open-source AI has become a major lever for the broader tech industry, accelerating experimentation and lowering barriers for researchers and smaller teams. If a company the size of Meta is indicating that open-source frontier models are not dependable enough for its top-tier use cases, that could influence how other large firms think about where to draw the line between public releases and fully controlled model stacks.

The comments arrive in a period when frontier AI development is moving quickly, and model choice is increasingly linked to product reliability, compute planning, and risk management. Large model performance is not just about benchmarks, it is also about how systems behave across real-world inputs, how quickly they can be audited, and how consistently they can be maintained.

Still, what Meta is not disclosing in the reported account limits how precisely the market can interpret the message. The remarks do not state whether Meta is concerned about specific open-source models, whether the issue is safety versus performance, or whether the conclusion is based on Meta’s internal evaluations of particular releases. There is also no detail on timelines for when Meta might change its stance.

Going forward, investors and developers will likely look for more specificity from Meta, including how it describes the criteria it uses to decide between open-source and in-house development, and whether it provides updates on its progress with internally built frontier models. Meta’s next product or research communications may also clarify whether “risk” refers to governance, technical robustness, or operational reliability, and how that affects partnerships and model distribution.