Ever since Meta announced its quarterly earnings, the upcoming foundational model, codenamed “Avocado,” has got all the hype. Now, according to an exclusive report by The Information, Avocado has completed its pre-training phase and is already outperforming leading open-source pre-trained models in testing. That’s according to an internal memo shared by Megan Fu, Product Manager of Meta’s Super Intelligence Lab.
Why Avocado’s early lests have everyone buzzing inside Meta
The memo, sent out last month, on January 20, suggests that Avocado, even before post-training optimization, can “compete” with top post-training models across knowledge, visual perception, and multilingual tasks. For those unaware, pre-training is the initial phase in AI development, where the system absorbs general knowledge and patterns from massive datasets.
Meta reportedly describes Avocado as the company’s “most powerful pretrained foundational model to date,” hinting at a major step forward to boosting AI goals this year. Not to forget, Avocado comes from Meta’s Super Intelligence Lab, which is a dedicated team overseeing the company’s AI research and next-generation projects.
While internal testing suggests that Avocado is ahead of open-source alternatives, independent verification remains unavailable at the time The Information published its report. Meta has yet to release the model publicly, meaning its real-world performance can only be assessed once external researchers and developers get access.
No concreate release timeline yet
Meta has not disclosed a timeline for Avocado’s public release or whether the model will be made open-source or restricted for internal use. However, Mark Zuckerberg teased last month that new models and AI tools will be released throughout the year in stages. So, we are assuming it’ll be out for the public. Given Meta’s rough run with the LLaMA model, Avocado has given new hope to the social media giant, which also happens to be a key contender in the ongoing AI race
