What made Meta hit the brakes on the Avocado launch
Meta made a move that few saw coming in the months ahead. Mark Zuckerberg’s company decided to delay the launch of Avocado, the artificial intelligence model that had been brewing behind the scenes with massive expectations. The main reason is directly tied to the model’s performance, which simply didn’t hit the bar needed to go toe-to-toe with what’s already out there. In a race where every technical detail matters, shipping something below the level of the competition would be a serious blow to the company’s reputation in the AI space.
Internal tests run by Meta’s research team evaluated Avocado across three areas considered essential for any cutting-edge AI model: logical reasoning, code generation, and writing quality. In all three categories, the model fell behind the latest competitors developed by Google, OpenAI, and Anthropic. We’re not talking about a subtle or marginal gap here. The results showed consistent shortcomings that made it clear more development time was needed before any public unveiling.
For a company pouring astronomical amounts of money into AI infrastructure, delivering a product that can’t hold its own against the competition would be an extremely tough situation to manage.
The timeline that slipped
The original plan was for Avocado to be finished and ready to integrate into Meta’s products by March 2026. The team at the internal lab called TBD Lab wrapped up the first phase of development, known as pre-training, at the end of last year. In January, they kicked off the next stage, post-training, and that’s when the mid-March launch target was set.
Now, with the official delay, the new launch window has been pushed back to at least May of the same year. That two-month gap might seem short at first glance, but in the world of generative AI, where new models and updates drop almost every week, every single day counts. The company needs to use that extra time not just to fix the issues flagged in internal benchmarks, but also to make sure the model hits the market with a clear, tangible edge that justifies all the hype.
One important thing to note is that Avocado isn’t a total failure. Based on the available information, the model outperformed Meta’s previous model and also did better than Google’s Gemini 2.5, which launched in March. However, it couldn’t keep up with Gemini 3.0, which arrived in November. So the model is making progress, just not fast enough to reach the front of the pack in the global AI race.
Billions on the line and the pressure from competitors
To understand how much weight this decision carries, it helps to look at the numbers in play. Last July, Zuckerberg publicly stated that Meta’s new AI models would push the technological frontier by about a year. Now, that promise is looking increasingly difficult to deliver on within the original timeframe, according to people familiar with the matter.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
Mark Zuckerberg projected spending up to $135 billion this year alone to keep Meta competitive in the global AI race. That figure is nearly double the $72 billion spent the previous year. The total covers everything from building and expanding data centers to acquiring cutting-edge chips and hiring elite researchers. The company also invested $14.3 billion in the startup Scale AI last June and brought its CEO, Alexandr Wang, age 29, on board as Meta’s head of AI.
It’s an investment that puts the company among the biggest corporate bets ever made in the tech industry. But even with that massive budget, money alone doesn’t guarantee that a company’s AI models will lead the performance rankings. The Avocado situation is the latest proof of that.
The rivals aren’t standing still
Competition in this space is getting fiercer by the day and shows no signs of slowing down. Google has been evolving Gemini with frequent updates that improve both reasoning capabilities and integration with its search and productivity products. OpenAI keeps pushing the boundaries of what its models can do, while Anthropic has carved out significant ground in the enterprise market with Claude, betting heavily on safety and reliability.
In this context, Meta can’t afford to release something that feels like it’s a generation behind. Public perception and the trust of developers who use the company’s open source models are directly tied to the quality of what gets shipped, and a premature launch could damage both in a lasting way.
Licensing Google’s Gemini? Yes, that was actually discussed
One detail that really turned heads behind the scenes is that Meta internally discussed the possibility of licensing Google’s Gemini as a temporary solution for its products while Avocado wasn’t ready. No decision has been made on that so far, but the mere existence of that conversation reveals just how urgent and tense things are inside the company.
Considering the use of a direct rival’s technology, even on a temporary basis, shows that the company’s leadership is well aware that its users and partners expect top-tier AI features integrated into the Meta ecosystem. Going without a competitive model simply isn’t a viable option when billions of people use WhatsApp, Instagram, and Facebook every day and expectations for smart features just keep growing.
The TBD Lab, internal tensions, and the open source question
The Avocado delay also brought to the surface some tensions that had been building within Meta’s AI team. The TBD Lab, assembled by Alexandr Wang, is an elite lab with around 100 employees working on two new models with fruit-themed names: Avocado, focused on language and reasoning, and Mango, geared toward image and video generation.
So far, the new AI division has shipped just one product, Vibes, an AI-generated video app similar to OpenAI’s Sora. The TBD Lab has also seen some turnover, with a few researchers leaving the team before Avocado even launched.
Clashes between leadership
Reports indicate that Wang had friction with Chris Cox, Meta’s chief product officer, and with Andrew Bosworth, the chief technology officer, over how the new AI models should improve the company’s advertising business. Building a large-scale language model is a process that involves hundreds of engineers and researchers working in sync, and when results don’t come as planned, pressure builds from every direction.
Last week, Meta announced to employees the creation of a new applied AI engineering team under Bosworth’s leadership, which will collaborate with Wang and the AI division. The move sparked rumors that Zuckerberg and Wang were at odds. The company moved quickly to shut that down, with a spokesperson calling the idea completely false. To drive the point home, Zuckerberg posted a selfie with Wang on Threads with the caption Meanwhile at Meta HQ.
Open source or closed source?
Another important debate surrounding Avocado’s development is the open source question. Meta has historically championed opening up its AI models, arguing that it helps advance the technology as a whole. The Llama family of models is the most iconic example of that philosophy, having built a huge developer community around the world.
However, over the course of last summer, Zuckerberg and Wang started considering keeping the new model closed source. Companies like OpenAI and Anthropic have long argued that letting others build on top of their AI models can pose security risks. If Meta goes down that path, it would mark a significant shift in the company’s strategy and could spark pushback from the developer community that already relies on the company’s open tools.
Adjusted expectations and what comes next
Even with the delay, Meta’s official stance is one of cautious optimism. In January, during an investor call, Zuckerberg adjusted his tone and said he expected the first models to be good, but that the most important thing would be to demonstrate the fast trajectory of progress the company is on.
Meta spokesperson Dave Arnold reinforced that message in a recent statement, saying the next model will be good, but that the focus is on showing the company’s speed of progress. He also emphasized that Meta will push the frontier consistently throughout the year, with continuous launches of new models.
AI experts who weighed in on the matter agree that there’s still time for Meta to catch up with its rivals. Improving artificial intelligence models is an iterative process, and occasional delays don’t necessarily mean the company will fall behind permanently. The challenge, though, is that every month without a competitive model on the market is a month where competitors solidify their positions and attract more developers and enterprise partners.
Llama 4, the lesson from the past, and the bet on the future
It’s worth remembering that this isn’t the first time Meta has struggled with its AI models. Llama 4, the company’s previous model, also fell short of expectations last year. It was precisely that experience that pushed Zuckerberg to invest heavily in Scale AI and restructure the AI division with Alexandr Wang’s arrival. Hiring top-tier researchers was also part of that recovery effort, with the company treating the best AI talent almost like professional sports stars.
Zuckerberg went so far as to declare that the company’s new goal is to create a superintelligent form of AI that would usher humanity into a new era. That’s a massive ambition that requires real results to be taken seriously by the market. And speaking of ambition, Meta is already thinking big for its next models. The successor to Avocado will go by the codename Watermelon, keeping the fruit-naming tradition alive, with the size of the name matching the scale of the ambition 😄.
The bigger picture in the AI race
This whole episode reinforces a truth that the artificial intelligence market has been learning the hard way: investing billions is necessary, but not enough. The AI race is defined as much by technical execution capability as it is by the speed of iteration and the quality of strategic decisions. It doesn’t matter if you have the best data centers or the most advanced chips if the training, fine-tuning, and evaluation process isn’t running in an integrated and efficient way.
Meta has the financial resources, the infrastructure, and the human talent to compete at the highest level. Now, with Avocado getting more time in the oven, the big question is whether the company can turn this delay into an advantage, delivering a model that doesn’t just match but surpasses what Google, OpenAI, and Anthropic are offering today.
The multi-billion-dollar battle for AI leadership is more alive than ever, and the coming months are shaping up to be decisive in determining who will truly lead this new era of technology. For anyone following the space closely, one thing is clear: the Avocado story is far from over, and what comes next could redefine Meta’s position in this global race 🚀.
