Meta Platforms has postponed the release of its flagship AI model, “Behemoth,” due to internal concerns over its performance improvements compared to previous models. Originally slated for an April 2025 launch to coincide with Meta’s first AI developer conference, the release was first delayed to June and is now expected in the fall or later .
Engineers have reportedly struggled to achieve significant advancements in Behemoth’s capabilities, leading to internal debates about whether the improvements justify a public release. This has caused frustration among senior executives, with potential management changes being considered within the AI product group responsible for the model’s development .
Despite public claims that Behemoth outperforms rival models from OpenAI, Google, and Anthropic on certain benchmarks, internal training setbacks have raised questions about these assertions . Additionally, Meta faced ethical concerns when it submitted an optimized version of a model for performance benchmarking rather than the publicly released version, leading to transparency issues .
The delay reflects broader challenges in the AI industry, where companies are encountering diminishing returns from scaling up models. Meta’s substantial investment in AI, with plans to spend up to $72 billion in capital expenditures this year, underscores the high stakes involved .AxiosInvestopedia