Meta has pushed back the release of its largest AI model, codenamed "Behemoth," delaying the rollout indefinitely amid internal doubts about its capabilities and mounting tensions within the company.
The massive language model was originally slated to launch in April to coincide with Meta's first developer conference, LlamaCon. Instead, Meta has only released two smaller models from the Llama family so far. The Behemoth launch was first moved to June and has now been pushed to fall or later, according to reporting from the Wall Street Journal.
People familiar with the project say Meta's engineering teams have struggled to deliver meaningful improvements over earlier versions. As a result, there's growing skepticism inside the company about whether the new model is ready for public release. Some developers are concerned that Behemoth's real-world performance won't live up to Meta's public claims about its superiority over competing models from OpenAI, Google, or Anthropic.
Has Meta missed a paradigm shift?
According to the report, Meta's leadership is increasingly frustrated with the team behind the Llama 4 models and is considering a shake-up within its AI product group. A Meta spokesperson declined to comment to the Wall Street Journal.
The company has faced setbacks before. The two Llama models released in April initially performed well on a popular AI chatbot leaderboard, but it later emerged that Meta had submitted a specially tuned version for the benchmark—not the public release. Mark Zuckerberg admitted that the benchmarked model had been optimized specifically for the test.
Meta's challenges aren't unique. OpenAI originally planned to launch GPT-5 by mid-2024, but in February, CEO Sam Altman announced that GPT-4.5 would come first, with GPT-5 expected later in the year, and that it would break from the traditional large language model format. Anthropic has also hit roadblocks: its larger model, Claude 3.5 Opus, was never released. Instead, both companies—along with Google and xAI—are now focusing on hybrid reasoning models, which have shown clear progress over their predecessors in multiple benchmarks. So far, Meta appears to have missed this shift in approach.
