11 out of 14 key scientists left the Llama team

Meta is going through a difficult period in the development of its artificial intelligence technologies. After the unsuccessful release of the Llama 4 model, the company faced new problems. The release of the Behemoth model has been postponed, which was supposed to become the most powerful in the Llama 4 lineup and was planned for release this month. Now the release dates are unknown, and it’s unclear whether this model will see the light at all.

There are several reasons for such a course of events. First of all, Meta’s management expresses serious dissatisfaction with the work of the team responsible for developing Llama. And plans significant changes in its composition. The situation is worsened by the fact that the team has already lost a significant part of key specialists. 11 out of 14 leading scientists who worked on early versions of Llama have left the company.

The second serious problem is related to the Behemoth model itself. Mark Zuckerberg previously made loud statements, promising that the new model would surpass leading language models from competitors — OpenAI, Google, and Anthropic. However, unforeseen difficulties arose during the training process, and the model shows results much worse than expected, which calls into question the advisability of its release in the current state.

These failures occur against the backdrop of a recent scandal when the company was accused of dishonest behavior when testing its models. Meta presented on the LMSYS platform not a user version of the model, but a specially optimized internal version. Tuned for specific tests of this platform, which is nothing other than manipulation of ratings and deception of users.