Meta memo claims Avocado AI better performing than Llama 4 models
Just months after the rollout of Llama 4, Meta is already preparing to move past it. According to a recent report by The Information, an internal memo from Meta’s elite Superintelligence Labs (MSL) reveals that the company’s next-generation AI model, codenamed "Avocado," is significantly outperforming the Llama 4 series in both power and efficiency.
The memo, authored by MSL Product Manager Megan Fu, describes Avocado as Meta’s "most capable pre-trained base model to date." This internal update signals a massive shift in Meta’s AI strategy as it looks to reclaim the lead from rivals like OpenAI and Google.
A "10x win" over Llama 4
The most striking detail in the leaked memo is Avocado's efficiency. Internal testing shows that Avocado achieved 10x compute efficiency "wins" on text-based tasks compared to Llama 4 Maverick, which was previously Meta’s flagship high-performance model.
Even more impressive is the comparison to the delayed Llama 4 Behemoth. The memo claims Avocado is over 100 times more efficient than Behemoth, suggesting that Meta’s new Superintelligence Labs has found a way to deliver "frontier-level" performance without the massive compute costs that plagued earlier projects.
Superintelligence labs: The team behind the pivot
The development of Avocado is the first major milestone for Meta Superintelligence Labs, a dedicated unit formed by Mark Zuckerberg in 2025. Led by Alexandr Wang (founder of Scale AI), the lab was created to move Meta beyond the "Llama era" and toward AGI-level reasoning.
While Llama 4 focused on native multimodality and "Mixture of Experts" (MoE) architecture, Avocado appears to be a ground-up rebuild. According to the memo, Avocado has completed pre-training. It is already outperforming leading open-source models in knowledge and visual perception. Despite being a "base" model (not yet fine-tuned for specific tasks), it is matching the performance of models that have already undergone extensive optimization.
The shift to "closed" AI?
For years, Meta has been the champion of open-source AI with the Llama series. However, the Avocado memo reinforces rumors that Meta is pivoting toward a proprietary or "closed" model strategy.
By keeping Avocado’s weights and architecture private, Meta could monetize the model through high-end enterprise tools and better protect its competitive edge. This shift reportedly contributed to the departure of Meta’s long-time AI chief, Yann LeCun, who remains a staunch advocate for open-source development.
What’s next for Avocado?
While Avocado is currently "finished" in its pre-training phase, it still needs to go through post-training and safety alignment before it’s ready for the public.
Meta is aiming for a first-half 2026 launch (likely Q1 or early Spring). Avocado (text/code) is being developed alongside "Mango," a next-gen model focused on high-fidelity image and video generation. As Meta prepares for its next major release, one thing is clear: the Llama family is no longer the top priority. The future of Meta AI belongs to the "Superintelligence" projects.