According to an internal memo obtained by "Information" magazine, Meta Platforms is increasingly optimistic about the first core large model that its newly formed AI team is about to launch. The memo, dated January 20, was released by Megan Fu, product manager of Meta Superintelligence Labs, which clearly states that the next-generation large language model Avocado has completed pre-training and is “Meta’s most powerful pre-trained basic model to date.” Pre-training is the initial stage of AI model development. The model needs to be exposed to massive data and learn general knowledge, patterns and relationships.
On the left are Alexander Wang and Mark Zuckerberg
The memo states that Avocado’s performance exceeds the current best open source base model. (The basic model refers to a large language model that has only completed pre-training and has not yet been optimized through the "post-training" stage; the post-training stage uses techniques such as human feedback reinforcement learning to fine-tune the model for specific tasks.) Although Avocado has not yet completed post-training, its knowledge reserve, visual perception, and multi-language capabilities are on par with the top post-training models.
Before Avocado is officially released to the public, it is still unclear whether Meta’s internal optimistic assessment of the progress of its AI model can withstand external inspection. But given Meta's rocky experience with AI over the past year, even overpromoting this progress internally would be risky for the company.
In 2025, the performance of Meta's previous generation large language model Llama 4 did not meet expectations, forcing the company to postpone its release. Meta eventually launched two versions of Llama 4 - Maverick (Ranger) and Scout (Scout), but some developers were disappointed with its actual performance.
These setbacks directly drove a large-scale restructuring of Meta’s overall AI business. Reorganization includes: Meta acquired Scale AI for US$14.3 billion in June last year. Alexandr Wang, CEO of the startup, subsequently joined Meta and was responsible for leading the new AI comprehensive organization-Meta Super Intelligence Laboratory. Since then, Meta has invested a record amount of money in the AI business: the company recently predicted that capital expenditures in 2026, including the cost of AI computing power, will increase by approximately 73% compared with 2025, reaching US$115 billion to US$135 billion.
Meta’s technological progress on new models may help control related costs. Another memo in mid-December last year showed that in text-related tasks, Avocado’s computing efficiency is 10 times higher than Maverick and more than 100 times higher than Behemoth. Behemoth is Meta's delayed, as-of-yet unannounced version of Llama 4 last year.
Meta said it successfully achieved efficiency breakthroughs by acquiring higher quality data, investing in model infrastructure, and adopting deterministic training, which ensures that the model outputs stable results under the same training method. This efficiency improvement helps reduce the energy consumption and cost of AI research and development - a crucial factor as Meta strives to catch up with competitors.
The positive tone of Meta's internal memo echoes recent public statements from executives. Last month, Meta Chief Technology Officer Andrew Bosworth hinted at similar efficiency improvements at a press conference at the World Economic Forum in Davos, Switzerland, saying that Meta's AI model "performed extremely well." Bosworth also said that developing a usable model for consumers requires a lot of post-training work.
At the same time, in last week's earnings call, Meta CEO Mark Zuckerberg said about the first large models planned to be launched by Meta's Super Intelligence Laboratory: "I expect our first batch of models will perform well, but more importantly, they will demonstrate our rapid development trajectory. I also expect that this year we will continue to launch new models and steadily advance the technological frontier."