In This Story
Meta (META) reportedly has delayed the release of its latest artificial intelligence model, known internally as “Behemoth,” raising concerns among employees about the effectiveness and direction of the company’s massive AI investments.
Engineers have struggled to improve the large-language model’s performance compared to earlier versions in the Llama model family, The Wall Street Journal (NWSA) reports. This has prompted internal debate about whether the upgrade is substantial enough to merit public release, sources told the Journal.
Shares in Meta were down more than 2% by Thursday’s close.
Initially set for an April debut during Meta’s first AI developer conference, Behemoth’s launch was pushed to June and potentially has been postponed until at least the fall.
When Meta released two smaller Llama models in April, the company was praised for the speed with which it has caught up with rivals when it comes to developing its own generative AI tools.
The company now faces increasing pressure to demonstrate meaningful progress with Behemoth and signal to Wall Street it can follow through with a return-on-investment. Meta has spent billions on AI development, with plans for up to $65 billion in capital expenditures this year. Its stock price to earnings ratio is 25.05, as of May 15, meaning investors are paying $25.05 for every $1 of earnings.
Though Meta has publicly claimed Behemoth outperforms competing models such as OpenAI, Google (GOOGL), and Anthropic in some benchmarks, internal sources told the Journal that the model is facing significant training challenges, and its real-world performance may not live up to the company’s public claims.
Meta’s AI journey began with its Fundamental AI Research Team, which developed the original Llama models in early 2023. However, 11 of the 14 authors of the initial Llama paper have since left the company, and newer versions are being built by a different team.
Controversy also arose in April when it was revealed that the model Meta submitted to a widely used AI chatbot leaderboard was not the same version released to the public. Meta later admitted it had optimized the submitted model specifically for the benchmark.
Meta is not alone in facing delays, however. OpenAI’s next major model, GPT-5, has also been postponed, with the company instead rolling out an intermediate version called GPT-4.5. Anthropic has likewise delayed the release of Claude 3.5 Opus, a larger and more powerful AI model, though it said the release is imminent.