Meta is holding back the release of a next-generation AI model codenamed 'Avocado' after it failed to consistently hit internal quality benchmarks, the New York Times reported on March 12. The company has not given a revised release date.
The delay is awkward timing for Meta. CEO Mark Zuckerberg has staked a significant part of his long-term strategy on AI leadership, and the company has spent heavily to get there — custom silicon through its MTIA chip program, a vast GPU buildout, and a recruiting push for AI research talent. The goal has been to run with OpenAI, Google DeepMind, and Anthropic. A stumble on a flagship model, even a temporary one, matters.
'Avocado' is believed to be a major upgrade within the Llama family, the model line that powers both Meta AI — embedded across Facebook, Instagram, WhatsApp, and Messenger — and Meta's open-source AI strategy. That second part is where the delay gets complicated. Developers and enterprises that build on Llama have come to depend on a predictable release cadence. When Meta slips, the ripple goes beyond its own products.
Internal delays are not rare in this industry — OpenAI and Google have both sat on models past intended timelines — but Meta is at a particular moment. It has been pushing Llama as the open-weight standard for the agentic era, and a public stumble hands competitors a talking point it would rather not give them. How fast it can close the gap on those benchmarks will say something about where the company actually stands.