Media reports indicate that since December of last year, Meta has been in the process of developing a text-based artificial intelligence model, internally codenamed 'Avocado'. Initially slated for release in the first quarter of 2026, the rollout has now been rescheduled to May. Despite several months of development, the model fell short in internal assessments of reasoning, coding, and writing capabilities when compared to top-tier models from competitors like Google, OpenAI, and Anthropic.

Nonetheless, Meta maintains that 'Avocado' represents its most advanced pre-trained foundational model to date. The company asserts that, even without fine-tuning, the model's knowledge base, visual perception, and multilingual abilities are already on a par with those of the current leading fine-tuned models. Additionally, Meta highlights that the model's computational efficiency has seen a marked improvement over its predecessors. This progress, according to Meta, can be attributed to the use of higher-quality data, substantial investments in model infrastructure, and the implementation of a 'deterministic training' methodology.