Meta’s Secret AI Models: Why the Open-Source Giant Is Going Closed
For three years, Meta was the undisputed champion of open-source AI. The Llama model family: from Llama 1 in 2023 to Llama 3.1 in 2024 and Llama 4 in 2025: was the foundation of an entire ecosystem. Startups, researchers, and enterprises around the world built on Meta’s open models, and Mark Zuckerberg personally evangelized the open-source AI philosophy in blog posts, interviews, and shareholder letters. Then, in early 2026, reports emerged that Meta was developing its most powerful models under internal codenames: “Avocado” (a frontier LLM) and “Mango” (a multimedia generator): and that these models might not be released as open-source.
The reports, first surfaced by industry journalists and subsequently confirmed by sources within Meta, indicate that the company is moving toward a “hybrid” approach under Chief AI Officer Alexandr Wang. Smaller, general-purpose models would continue to be released openly, maintaining Meta’s developer ecosystem and community goodwill. But the most powerful frontier models: those with capabilities that raise safety concerns or represent significant competitive advantages: would remain proprietary, available only through Meta’s own products and APIs.
This represents a fundamental strategic reversal. Zuckerberg’s original argument for open-source AI was compellingly simple: Meta makes money from ads and social media, not from selling AI models. Open-sourcing models commoditizes the technology layer, attracts developers to Meta’s ecosystem, and delivers free improvements from the global research community. The strategy worked brilliantly: Llama became the most widely used open model family in the world.
Several factors appear to be driving Meta’s shift:
Safety concerns. As AI models become more capable, the risks of open release increase. A model that can generate highly persuasive misinformation, create realistic deepfakes, or assist with harmful activities poses risks that don’t exist with less capable models. The Anthropic Mythos situation: a model restricted because of its vulnerability-finding capabilities: illustrates that frontier capabilities can create genuine safety dilemmas.
Competitive intelligence. Open-sourcing a model means your competitors can study your architecture, training methodology, and capabilities in detail. With Chinese AI labs reportedly using open models as baselines for their own development (and allegedly using distillation to harvest intelligence from API-accessible models), the competitive cost of openness has increased.
Please enable JavaScript to read the full article.