Meta has delayed the release of its larger generative AI model, Llama 4 Behemoth, from its original April launch to an unspecified date in the fall. The model’s multiple delays, reported by The Wall Street Journal, come as critics inside and outside of Meta question whether large generative AI models have reached a performance plateau.
What is Llama 4 Behemoth?
Llama 4 Behemoth is a 288-billion-parameter large language model. Meta described Behemoth as “one of the smartest LLMs in the world and our most powerful yet to serve as a teacher for our new models.”
Originally, Meta planned to debut Behemoth at its AI developer conference in April. The release was first delayed until June, and has now been postponed again. Behemoth would be the largest version of Llama 4, Meta’s latest flagship model. Meta claimed Behemoth outperformed OpenAI’s GPT-4.5, Anthropic’s Claude Sonnet 3.7, and Google’s Gemini 2.0 Pro on several STEM benchmarks.
Meta has already used Behemoth to train its smaller Llama 4 models, Scout and Maverick.
Meta competes with OpenAI, Google, Anthropic, and xAI, and other companies in the generative AI market.
SEE: xAI’s Groq infrastructure enables high-speed output for Meta’s Llama API.
Doubts raised about AI performance leaps
According to The Wall Street Journal, some Meta employees are questioning whether Behemoth offers a significant enough improvement over predecessors to warrant a public release. At the same time, senior executives are blaming the Llama 4 team for not making enough progress.
These internal concerns echo broader doubts within the AI industry about the pace and cost of advancing generative AI. Some experts warn that further gains may come with prohibitively high costs and slower development cycles, making it difficult to sustain the rapid pace of product launches seen in recent years by companies like Meta and Open AI.
“Right now, the progress is quite small across all the labs, all the models,” Ravid Shwartz-Ziv, an assistant professor and faculty fellow at New York University’s Center for Data Science, told The Wall Street Journal.
