Meta plans partial open‑source push

Meta is not abandoning its open model strategy. It is narrowing it. According to an Axios scoop published Monday, the company is preparing to release open versions of its next AI models, but only after a safety review, and not with every piece of the system exposed. The shift matters because Meta has spent the past three years casting itself as the big tech company willing to put powerful models directly in developers’ hands. Now it is drawing a line between the parts it wants widely used and the parts it wants to keep. (theverge.com) That is a meaningful change, not a cosmetic one. In Meta’s earlier Llama releases, the pitch was simple: download the model, run it yourself, fine-tune it, and avoid paying a frontier lab every time a user sends a prompt. That promise made Meta unusually important to startups and independent developers, especially compared with OpenAI, Anthropic, and Google, which mostly sell access through hosted APIs. Meta still seems to want the political and ecosystem benefits of being the company that “opens” models. It just no longer wants to give away the whole machine. (ai.meta.com) The timing explains a lot. Meta’s next major foundation model, code-named Avocado, was reported in March to have slipped after internal tests showed it lagging the strongest systems from Google, OpenAI, and Anthropic on reasoning, coding, and writing. The same report said Avocado beat Meta’s previous model and even topped Google’s Gemini 2.5, but not Gemini 3.0, which was enough to force a delay to at least May. A company that is behind at the frontier has less incentive to publish every technical advantage it still has. (businesstimes.com.sg) That pressure collides with another problem: Meta’s models have never been as open as the marketing suggested. The Llama 4 license, for example, is a custom community license, not a standard open-source license. It allows broad use and modification, but it also imposes Meta’s own terms, including branding requirements and a special permission threshold for companies above 700 million monthly active users. Even the definition of what counts as “open source AI” has become more contested, with the Open Source Initiative arguing that systems like Llama do not meet the standard in the strict sense. So the new plan is less a betrayal of purity than an admission of reality. (github.com) What Meta appears to be building instead is a hybrid stack. Developers may get model weights or deployable versions they can run on their own infrastructure, while Meta keeps some training methods, alignment layers, or other high-value components private. Safety vetting before release adds another filter. The Verge’s account of the Axios report says Meta wants to make sure the models do not introduce “new levels of safety risk” before wider release. That language suggests the company is thinking less like an open-science lab and more like a platform owner deciding which capabilities are safe to export. (theverge.com) For startups, this creates a middle path that is useful precisely because it is messy. A fully closed API can be expensive and limiting. A fully open stack can be hard to support and risky to govern. If Meta offers partially open models with enough freedom to self-host and customize, many companies will take that trade. But they will need to read the license and release notes with unusual care, because “open” may describe the weights, not the training data, not the safety tooling, and not the most strategically important components. Meta’s last big open-model push helped define the market. Its next one may define how much openness is left when the race gets serious.