Meta, which has long been one of the few leading AI companies to release open-source large language models (LLMs), may be shifting its strategy toward closed-source models, The New York Times reports.
According to sources, at the recently created Meta Superintelligence laboratory, several leading experts, including Alexander Wang from ScaleAI, discussed the possibility of abandoning the open Behemoth model in favor of creating a closed version.
Meta had planned to release Behemoth months ago and even finished training it, but the release has been repeatedly delayed due to concerns about its performance. As reported in May, engineers doubt that the model's leap in capabilities over previous versions is significant enough for release.
The move to closed models will require Meta to make not only technical changes but also a change in approach within the company. Meta has been praised for its openness in the past, publishing the code for its AI models, which has helped the industry develop. In particular, Chinese startup DeepSeek is partly based on its developments. One of the company’s top executives, Yann LeCun, has also emphasized that an open platform will win in the field of AI.
In turn, Mark Zuckerberg, founder and CEO of Meta, said during a podcast last year that his company is "obviously very supportive of open source," but that doesn't mean it will open source every single one of its products.
Discussions at Meta Superintelligence are still ongoing, and nothing has been proposed to Zuckerberg, who must finally approve any changes.
Recall that in addition to creating a new laboratory for the development of superintelligence, Meta is also creating supercomputers with a capacity of over 1 GW for its further development.