Meta's flagship AI model Behemoth delayed release raises market concerns
Meta Platforms has delayed the release of its flagship AI model "Behemoth" to the fall due to development setbacks, triggering a crisis of market confidence. This model employs a mixture of experts architecture aimed at enhancing computational efficiency but faces technical bottlenecks. Meanwhile, Meta must contend with competitive pressure from Chinese tech companies and balance technological breakthroughs with cost control. This delay reflects the common challenges faced by large tech companies in the AI field
According to Zhitong Finance APP, social media giant Meta Platforms (META.US) is facing a crisis of market confidence due to setbacks in the development of its flagship artificial intelligence model "Behemoth." Internal sources indicate that the company's engineers have encountered technical bottlenecks while optimizing this ultra-large language model with 2 trillion parameters, causing the originally scheduled release in June to be postponed until autumn or even later.
As the core of Meta's new generation AI strategy, the "Behemoth" in the Llama 4 series of models was highly anticipated. This model employs an industry-leading mixture of experts (MoE) architecture, significantly enhancing computational efficiency through task modularization. Currently, derivative versions based on its architecture, Maverick and Scout, have been opened to developers, allowing users to experience its native multimodal capabilities through Meta's applications or official website—this model can simultaneously process multimedia inputs such as text, images, and videos, and has outperformed several leading competitors in benchmark tests.
Market analysis suggests that Meta's AI advancement is under dual pressure: on one hand, the race against time between American tech companies and their Chinese counterparts is intensifying. Since the release of DeepSeek earlier this year, Chinese firms such as Alibaba (BABA.US), Tencent (00700), and Baidu (BIDU.US) have accelerated their technological iterations, with Baidu even making its ERNIE Bot service completely free, while Tencent has deeply integrated AI features into the WeChat ecosystem; on the other hand, Meta itself must balance technological breakthroughs with cost control—its MoE architecture is modeled after the successful experience of DeepSeek in reducing model operational costs.
It is noteworthy that this delay exposes the common dilemma faced by large tech companies in the AI arms race. Although Meta made a high-profile release of the Llama 4 series in April, the commercialization of the core model still needs to overcome engineering challenges. According to insiders, the R&D team is focusing on improving model stability and output consistency, two metrics that did not meet expected standards in preliminary tests.
Industry observers point out that in the critical phase of generative AI transitioning from the laboratory to industrial applications, the competition between technological maturity and engineering capability may determine the direction of the next stage of the global AI race. Whether Meta can maintain technological leadership through architectural innovation will become an important indicator of its future valuation
