Meta plans partial open‑sourcing

Meta is not abandoning open AI. It is narrowing it. Axios reported on April 6 that the company plans to release open versions of its next AI models, but only after first keeping some of the strongest systems proprietary. These would be the first new models built under Alexandr Wang, the former Scale AI chief now leading Meta’s AI push. The shift matters because Meta spent the last three years presenting openness as both principle and strategy, and now it is drawing a line between what it will share and what it will keep for itself (axios.com, reuters.com). That line did not come out of nowhere. Meta built real momentum with Llama. In March 2025, the company said Llama had passed 1 billion downloads, a number it used to argue that open models were becoming a serious alternative to closed ones. A month later, Meta released Llama 4 Scout and Maverick as open-weight models and described them as the start of a new multimodal generation. It also said a larger model, Llama 4 Behemoth, was still in training and not yet being released. Even in that announcement, you could see the split forming: some models were for the ecosystem, and some were still being held back (about.fb.com, about.fb.com). The company’s own language makes the logic plain. Meta still says open access helps developers with transparency, customization, and security, and its responsible-use guides are built around the idea that outside builders will add their own safeguards, policies, and product controls on top of foundation models. But those same documents also show why a hybrid strategy is attractive. The hard part is no longer just training a model. The hard part is turning it into a usable system with input filters, output checks, red-teaming, reporting tools, and domain-specific controls. If Meta keeps the sharpest model private while releasing a trimmed or delayed version, it can still claim the ecosystem benefits of openness while protecting the pieces it thinks are easiest for rivals to copy (ai.meta.com, ai.meta.com). This is also a response to pressure. Reuters reported in June 2025 that Zuckerberg reorganized the company’s AI work under Meta Superintelligence Labs after staff departures and a weak reception for Llama 4, while rivals such as OpenAI, Google, and DeepSeek gained momentum. Meta then put Alexandr Wang in charge of the new division and tied the AI effort more tightly to products that can make money, including the Meta AI app, ad tools, and smart glasses. Once the goal becomes product advantage instead of pure ecosystem influence, full openness gets harder to justify. You do not spend billions on talent and infrastructure just to hand every frontier trick to competitors on day one (reuters.com, about.fb.com). That is why the most practical consequence may land far below the level of frontier research. A partially open Meta means more work for people who can package, govern, and deploy models rather than invent them. The company has already been building that layer through developer guides, grants, APIs, and product integrations around Llama. If the next wave arrives as a staggered release, with proprietary systems first and open variants later, the gap will be filled by documentation, evaluation harnesses, safety wrappers, fine-tuning workflows, and adoption plans inside companies that want the benefits of open models without the chaos. Meta’s strategy is becoming less ideological and more operational, which is another way of saying the interesting work is moving one layer down the stack (about.fb.com, about.fb.com).