Meta told employees this week that its next AI model has finished pretraining and is already beating leading open-source alternatives, according to an internal memo obtained by The Information. The model, codenamed Avocado, hasn't even started post-training yet.
The January 20 memo from Megan Fu, a product manager at Meta Superintelligence Labs, called Avocado the company's "most capable pre-trained base model to date." More interesting: it's apparently competitive with models that have already gone through extensive optimization, at least on knowledge, visual perception, and multilingual tasks.
What the efficiency claims actually mean
One report from Digit cites the memo claiming Avocado achieved "10x compute efficiency wins" on text tasks compared to Llama 4 Maverick, and "100 times more efficient" than the troubled Llama 4 Behemoth. Those numbers sound impressive until you remember that Behemoth was shelved for underperforming on benchmarks and never actually shipped. Comparing against your own failed project isn't exactly a high bar.
The comparison to open-source base models is the more telling detail. Base models, before post-training, are essentially raw intelligence without the polish that makes them useful for specific tasks. That Avocado is supposedly competitive with fully trained models while still in this raw state is either a significant achievement or the kind of claim that falls apart under independent testing.
The open source question
Here's where it gets messier. Bloomberg reported in December that Avocado may launch as a closed model, breaking from Meta's Llama tradition of releasing weights publicly. Chief AI Officer Alexandr Wang, who joined after Meta's $14.3 billion Scale AI deal, reportedly favors closed models.
Zuckerberg himself has shifted tone. After writing a memo titled "Open Source AI Is the Path Forward" and publicly saying "fuck that" about closed platforms, he's now talking about being "careful about what we choose to open source." The trigger, reportedly, was DeepSeek's R1 model incorporating Llama's architecture, which some at Meta found galling.
There's an irony here. Reports suggest Avocado's training incorporated multiple open models from competitors, including Alibaba's Qwen, Google's Gemma, and OpenAI's gpt-oss releases. Meta may be using the fruits of open-source development to build something it keeps closed.
What's next
Avocado is targeting Q1 2026 release, slipped from an original late-2025 goal. It's being developed inside TBD Lab, a smaller elite group within Meta Superintelligence Labs. A companion image and video model codenamed Mango is also in development, with both targeting first-half 2026 availability.
The pretraining milestone matters because it's the computationally expensive part. Post-training, which involves fine-tuning for specific tasks and safety alignment, typically takes less time. Whether that means Meta can hit its spring window depends on how smoothly that phase goes.
Independent benchmark results will tell us whether the internal hype matches reality. Until then, these are just claims in a memo.
