Meta delays flagship Behemoth model due to performance issues

The 2 trillion parameter model was teased at the launch of Llama 4 in early April, and was slated for a release at the LlamaCon, Meta’s AI developer conference in late April.

Now it seems following internal dismay at the model’s performance, it was delayed until June, and has just been postponed again until sometime in the fall or later, writes Reuters.

Internal frustration
Meta is investing multiple billions into its AI program, and senior executives are increasingly frustrated at the Llama 4 team, according to the Wall Street Journal. They are now eyeing management changes, the paper writes.

In April, Meta said of the model that it was «one of the smartest LLMs in the world and our most powerful yet to serve as a teacher for our new models»

Behemoth was supposed to significantly outperform current state of the art models on some tests, but in reality, it has faced seen performance hobbled by training challenges.

Brain drain abounds
The first versions of Llama received no small amount of praise, but they were built by another team at Meta, the Fundamental AI Research Team. This team was stacked with PhD level academics and researchers, and 11 of the 14 researchers have left the company since early 2023.

There was also an incident where Meta was accused of tricking the AI benchmark LMAarena with the Original Llama 4 Maverick model, reaching a second place on the leaderboard with a finely tuned model.

The production model was later added to the benchmark and placed 38th.

Read more: Paywalled WSJ, a writeup on Reuters, and discussion on r/singularity.