Meta's internal memo signals AI comeback after rocky year

Avocado outperforms the best freely available base models, according to the memo from Megan Fu, a product manager at Meta Superintelligence Labs. In knowledge, visual perception, and multilingual performance, it can even compete with leading fully trained models - even though Avocado hasn't gone through the second training phase yet. This phase, called post-training, refines models for specific tasks.

The model also uses computing power far more efficiently: ten times more efficient than Maverick and a hundred times more efficient than Behemoth, two earlier Meta models. Meta achieved these gains through better training data, improvements to the technical foundation, and changes to the training method.

Avocado could validate Zuckerberg's billion-dollar AI bet

These advances matter for Meta, which had significant problems with its Llama 4 model in 2025. The release was delayed multiple times, benchmarks were manipulated, and developers were disappointed with performance after launch. These struggles led to a major reorganization of the AI division, including Yann LeCun's departure. Meta invested $14.3 billion in the startup Scale AI and brought in its CEO Alexandr Wang to lead a new AI unit called Meta Superintelligence Labs.

Meta is now pouring record amounts into its AI efforts: for 2026, the company expects capital expenditures between $115 and $135 billion - about 73 percent more than 2025. CTO Andrew Bosworth recently said at the World Economic Forum in Davos that Meta's AI models are very good but still need significant post-training. CEO Mark Zuckerberg echoed this view at the last earnings call. Meta plans to release new models steadily throughout the year. According to rumors, Meta will move away from the open-source approach of its Llama models with Avocado. A model focused on visual applications with the codename Mango is also in development.