Meta to open-source models

Meta is preparing to release open-source versions of its next AI models, according to Axios, and that matters because Meta is the last big U.S. lab still treating frontier AI as something outside developers should be allowed to modify, run, and build on themselves. The report says these will be the first new models developed under Alexandr Wang, the Scale AI founder Meta brought in to lead its new AI push, and that the company plans to release them under an open-source license eventually, not necessarily on day one (axios.com, theverge.com). That word, “eventually,” is doing real work. The Verge reports that Meta wants to keep some pieces proprietary at first and make sure the models do not introduce fresh safety risks before wider release, which suggests a slower and more cautious version of the company’s old Llama playbook (theverge.com). The change is subtle but important. Meta is not walking away from openness. It is trying to preserve the political and commercial value of openness while giving itself more control over timing. That tension has been building for months. Meta spent the last two years arguing that open models are not a side project but a strategic bet. In July 2024, Mark Zuckerberg wrote that open-source AI was “the path forward” and compared it to Linux, arguing that developers wanted models they could fine-tune, distill, and run with lower cost and fewer restrictions than closed systems allow (about.fb.com). Long before that, when Meta launched Llama 2 in July 2023, it said the model was free for both research and commercial use and explicitly pitched open access as better for experimentation and even for safety, because more people could inspect and stress-test the system (about.fb.com). Meta has numbers now to show that this argument found an audience. In March 2025, the company said Llama had passed 1 billion downloads, and it used that milestone to highlight a broad range of users, from Spotify to startups and independent developers. Meta’s message was simple: open models create an ecosystem, and ecosystems are hard for rivals to copy once developers have built tools and businesses around them (about.fb.com). That is the real backdrop for the new report. Open source is not just branding for Meta. It is one of the few places where the company can still claim a distinct position in the AI race. The timing also says something about the market Meta is facing. OpenAI and Anthropic built their businesses around tightly controlled APIs. Google has mixed approaches, with both proprietary systems and more open releases. Meta chose the opposite flank. It gave away model weights, let companies host and adapt them, and tried to make “you control it” into a product feature. Axios reports that Meta does not expect its upcoming models to beat every rival in every category, which makes the open-source promise even more central. If you cannot win cleanly on raw benchmark prestige, you can still win by being the model developers are allowed to actually use (axios.com, about.fb.com). There is also a national angle, and Meta has been pushing it hard. In late 2024, the company argued that open-source AI could help the U.S. lead globally and said it was making Llama available to U.S. government agencies and contractors working on national security applications (about.fb.com). The Axios report fits neatly into that frame. A U.S.-based open model is not just a convenience for developers who want to run systems locally. It is also a way for Meta to say that American AI leadership should not belong only to companies that keep their best models behind a paywall. That is why this story is bigger than one product release. Meta appears to be trying to split the difference between two instincts that usually collide. It wants the leverage that comes from controlling a frontier model early, and it wants the reach that comes from letting everyone else tinker with it later. If the company follows through, developers may get what closed-model rivals still resist: new high-end models they can inspect, adapt, and deploy on their own terms, even if they have to wait a little longer than before for the weights to drop (axios.com, theverge.com).