Meta opens lower-tier AI models, keeps flagship private

Meta will publish a set of new AI models under the leadership of Alexandr Wang, making some lower- and mid-tier variants available under open-source licenses while keeping its most advanced systems private for the near term. The company plans to provide parts of the models to developers and consumers but will withhold key components and full training details.

Some variants will be licensed to allow modification and local use, while the core architecture and certain training data will remain closed. Meta says the partial openness is intended to reduce safety risks and protect proprietary capabilities while still enabling developer access.

Company officials describe the approach as a middle ground between broadly shared research models and fully closed commercial systems. Meta aims to attract developers who want modifiable, U.S.-based options and to preserve top-tier models that it can integrate into its products and services.

Meta plans to embed the new models into Facebook, Instagram and WhatsApp, giving billions of users access to AI features often without a direct charge. Executives expect that consumer reach through those platforms will be an advantage even if the company does not lead every technical benchmark.

The move marks a change from prior releases in which Meta allowed wider modification of its frontier models. Earlier Llama releases were notable for broad developer access; the Llama 4 family did not match some competitors on capability measures, and the company seeks to narrow that gap with the new lineup while limiting exposure of its top systems.

Other large technology firms have also restricted access to their most advanced models, citing safety and commercial concerns. Competing companies have focused on enterprise and government customers and have published fewer frontier models for public modification.

Debate over how much powerful AI should be openly available has increased. Some public figures have called for broader sharing of models, while others have warned about the risks of misuse and argued for tighter controls on distribution and deployment.

Meta plans to make lower- and mid-tier variants accessible to sustain developer engagement, while keeping flagship systems private to manage safety risks and protect competitive advantages in consumer products.