Meta's Alexandr Wang Is Preparing Its First AI Models—With a Hybrid Open-Source Twist

When Alexandr Wang joined Meta last year in a deal that gave the company access to Scale AI’s data infrastructure and Wang’s personal prestige, the arrangement was described publicly as something between a strategic acqui-hire and a bet on the next generation of AI leadership. Wang, who founded Scale AI at 19, sold it to Palantir, and co-founded Scale AI again before becoming its CEO, was 27 at the time of the deal and had already built one of the most valuable AI data companies in the world.

Now, according to reporting by Axios on April 6, the first tangible product of that arrangement is taking shape—and with it, a significant shift in Meta’s foundational approach to AI.

A Hybrid Model: Open Below, Proprietary Above

Meta’s plans, as described to Axios by people familiar with the matter, call for a two-tier release structure for the new model family. Smaller, less capable versions will be open-sourced under a permissive license, continuing Meta’s tradition of releasing models that developers can run, fine-tune, and deploy freely. The largest, most capable frontier models, however, will remain proprietary—available via API but not available for download or local deployment.

This is a meaningful departure from the Llama series, which Meta has released under a permissive license since 2023. Llama models became the backbone of a vast developer ecosystem: startups built products on them, researchers used them for studies that would be impossible with closed models, enterprises adopted them to avoid API dependency, and emerging-market developers in regions with limited access to expensive cloud APIs built on them because they had no other option.

That ecosystem was not accidental. Meta built it deliberately as a counterweight to OpenAI and Anthropic, which operate purely proprietary stacks. The Llama family gave Meta credibility with developers, reduced the gravitational pull of OpenAI’s API, and generated the kind of network effects that commercial strategies struggle to manufacture. Walking away from it, even partially, requires a compelling reason.

Why the Strategic Shift?

Two dynamics are pushing Meta toward a hybrid model. The first is competitive. OpenAI’s o3 and o4 series have established a substantial reasoning capability gap over fully open-weight models, and internal evaluations at Meta reportedly show that the models Wang’s team is building are good enough to compete at the frontier—but not if they are released for competitors to study, adapt, and distill from. The decision to keep the most capable models proprietary is, at its core, a decision that the moat of capability is now worth protecting.

The second dynamic is commercial. Meta has historically struggled to monetize its AI investments in a way that satisfies Wall Street. Releasing models for free generates good will and ecosystem density, but it does not generate revenue. With Meta’s Reality Labs continuing to lose billions per quarter and AI infrastructure costs scaling rapidly, the pressure to demonstrate that AI can be a revenue center—rather than a cost center with public relations benefits—is intensifying. A proprietary frontier model can be charged for. A fully open one cannot.

Wang has reportedly framed the hybrid strategy around a democratization narrative: the open-sourced smaller models will still give developers worldwide access to U.S.-built AI, while the proprietary larger models represent Meta’s competitive edge. Whether that framing holds up depends on exactly where in the capability curve the line is drawn.

The Developer Ecosystem Gamble

The risk is real. Meta’s reputation in the developer community rests substantially on its commitment to openness. The Llama series generated an enormous amount of goodwill and practical leverage: entire startups were built on the promise that Meta would not pull the rug. A shift to hybrid licensing would signal that the rug can, in fact, be pulled—just not immediately and not completely.

More practically, the transition to a hybrid model could fragment the ecosystem. Developers who have built workflows around local Llama deployment will face a harder choice: accept API dependency for the capabilities they need, or constrain themselves to the open-weight tier. For enterprise users who chose Llama precisely to avoid vendor lock-in, this changes the calculus substantially.

The competitive implications for Hugging Face, Mistral, and other open-weight AI labs are also significant. Meta’s full-openness policy functionally subsidized the entire open-source AI movement by providing a highly capable baseline that smaller open-weight labs could benchmark against and build upon. A Meta that no longer releases its best models open-weight removes a powerful force that had been keeping the open ecosystem competitive with proprietary systems.

Wang’s Vision for Meta’s AI Position

Wang, since joining Meta, has been explicit about his belief that the United States needs to win the AI race against China, and that democratizing access to U.S.-built AI tools is a national security argument as much as a commercial one. In public remarks, he has emphasized that Meta’s open models have been adopted in countries that would otherwise default to Chinese AI systems, providing an implicit counterweight to DeepSeek and other Chinese open-weight models that have gained significant global traction.

The hybrid strategy is consistent with that framing: keep giving the world access to competitive U.S.-built AI (via the open-weight smaller models), while reserving the capability ceiling for revenue generation and competitive protection. It is a more sophisticated argument than “we will be fully open forever,” but it is also a more fragile one—dependent on the smaller models remaining genuinely useful even as the frontier continues to advance.

What to Expect

Meta has not announced a release date for the new model family. The Axios report does not specify model sizes, training details, or capability benchmarks. What is clear is that Alexandr Wang’s arrival at Meta was always expected to produce something different from the Llama series, and the hybrid licensing structure represents the first evidence of what that difference will look like in practice.

For the open-source AI community, this is a moment to watch carefully. If Meta’s most capable models are proprietary, the open-weight frontier—which Llama, Mistral, and others have been advancing—falls further behind the commercial frontier than at any point in the past three years. Whether that gap can be closed by the open-weight community alone, without Meta’s frontier models as a reference point, is one of the defining questions of the next phase of AI development.