Meta's Avocado LLM Outperforms in Pre-Training Trials (META-US)

Meta Platforms (META-US)’s internally developed Avocado large language model delivered strong results in early pre-training trials, according to a January 20 memorandum from the Meta Superintelligence Labs. The memo states Avocado is now the company’s most powerful pre-trained foundation model, outperforming leading open-source counterparts in knowledge, visual understanding, and multilingual capabilities prior to post-training fine-tuning. The project is led by the newly formed Meta Superintelligence Labs, established under CEO Mark Zuckerberg’s restructuring of Meta’s AI initiatives. The team, headed by Scale AI’s Alexander Wang, is expected to release the model in the first half of 2026, though a formal timeline and open-sourcing plans have not yet been announced.