Meta has postponed the launch of Behemoth, the largest version of its open-source Llama 4 generative artificial intelligence model, until at least the fall. According to The Wall Street Journal, the delay stems from the model’s lack of ´significant´ performance improvements, prompting Meta to push its release from the previously expected summer timeframe. This setback marks the first major delay for Meta’s rapidly iterated Llama series, which has positioned itself as an open-source counterweight to closed models created by industry leaders such as OpenAI, Google, and Amazon.
Industry observers note that the business impact of Behemoth’s delay is relatively muted, as many companies already leverage existing open-source Llama models or opt for proprietary offerings from cloud providers. Smaller companies and research organizations benefit from open-sourced models like Llama but often face challenges deploying them at scale, especially since Meta does not provide enterprise deployment services. In the interim, Meta has released two substantial yet smaller Llama 4 models: Maverick, with 400 billion parameters and a 1 million token context window, and Scout, with 109 billion parameters and an even larger 10 million token context window. Behemoth, initially set to boast 2 trillion parameters, was intended to debut alongside these models.
The delay highlights deeper technical and strategic frustrations within Meta’s leadership. Internal sources report mounting impatience regarding the Llama team’s progress, with discussions of leadership changes on the table amid heavy capital expenditures earmarked for artificial intelligence development. A core challenge for Meta—and for the broader industry—is overcoming the diminishing returns from scaling laws and the scarcity of high-quality training data, as noted by similar delays at OpenAI with its own next-generation models. Leading artificial intelligence companies continue to lobby for expanded data access to sustain model improvements, but algorithmic and data constraints are increasingly apparent. Thus, Meta’s setback underscores the industry-wide hurdles facing generative artificial intelligence advancement, rather than being an isolated misstep.
