Meta: New AI models to be partly open-source

Meta plans to release parts of its upcoming AI models under open-source licenses. The company wants to make a significant portion of the source code of new models freely accessible, reports the US news platform Axios. However, Meta will keep some components of the code proprietary for security reasons.

Some of Meta's new AI models could thus go beyond the open-weight architecture of previous models. “Open-weight” refers to partially open AI models whose trained parameters (“weights”) are freely available, but not the code and training data. Meta has long relied on such partially open models for its AI models.

Hybrid AI Strategy

According to Axios, Meta plans a hybrid AI strategy overall: The most powerful of the upcoming models will likely be released as closed models, while others will be open-source. This could serve to attract AI developers while simultaneously securing potential competitive advantages of the largest models, Axios writes.

The new AI models would be the first to be developed under the leadership of Alexandr Wang. The new open-source strategy in AI development is to be significantly shaped by him. Wang wants to further open up Meta's AI technologies and thus make them more attractive to developers. In addition, Meta wants to focus more on end consumers – especially since the competition, primarily Anthropic and OpenAI, is increasingly focusing on business with large corporations and governments.

AI Offensive with Little Success

Wang was only appointed Meta's Chief AI Officer in the middle of last year. The 29-year-old is the co-founder and former CEO of Scale AI, a company that sells annotated training data for AI applications. In the summer of last year, Meta acquired 49 percent of Scale AI for around 14 billion US dollars, after which Wang moved to Meta.

The acquisition of Scale AI shares is part of a multi-billion dollar AI offensive running since 2025, with which Meta is specifically recruiting highly qualified AI specialists. The goal is to take a leading role in the competition for advanced AI models up to superintelligence. A central component of this strategy is the Meta Superintelligence Labs (MSL) unit, founded in the summer of 2025, whose head is Alexandr Wang since his move to Meta.

So far, however, Meta's AI models have lagged far behind the performance of the competing model families from Anthropic or OpenAI. The top model Llama 4, in particular, caused considerable problems. The release was postponed several times, the company manipulated benchmarks, and developers were disappointed with the performance. Meta subsequently began its recruitment offensive and restructured its AI department.

(rah)