Meta’s AI Strategy Just Got Weird: Open Source, Closed Source, and the Avocado Problem
Meta has been the loudest champion of open-source AI for two years. Llama became the foundation that thousands of companies built on. Zuckerberg wrote an entire letter about why open source is good for everyone.
And then they quietly started building a closed-source model called Avocado.
What’s going on?
The Llama Story So Far
Let’s give credit where it’s due. Meta’s Llama models changed the AI space. Before Llama, if you wanted a powerful language model, you had two options: pay OpenAI or pay Google. Llama gave everyone a third option: run it yourself.
Llama 3.1 pushed context length to 128K tokens. Llama 4 introduced natively multimodal models — Scout and Maverick — that could handle text, images, and video in a single architecture. At LlamaCon in April 2025, Meta went all-in on the “the future is open source” message.
And the impact was real. Startups built products on Llama that they couldn’t have afforded to build on proprietary APIs. Researchers used it to advance the field. Countries used it to build AI capabilities without depending on US tech companies.
So why would Meta start hedging?
Enter Avocado
Reports surfaced that Meta is developing a new model codenamed “Avocado” under its Meta Superintelligence Labs (MSL), led by Chief AI Officer Alexandr Wang. The key detail: Avocado is being developed under “tighter control” — which is corporate speak for “not fully open source.”
The original plan was to launch by end of 2025. That got delayed. Current estimates put it in Q1-Q2 2026.
Why the shift? A few reasons that make sense when you think about it:
Llama 4’s reception was lukewarm. Despite being technically impressive, Llama 4 didn’t generate the same excitement as Llama 3. The market is getting saturated with open models, and differentiation is harder.
Revenue pressure. Meta has spent billions on AI infrastructure. At some point, investors want to see returns. Open-sourcing your best models makes it hard to charge for them.
Competitive dynamics. OpenAI, Google, and Anthropic all keep their best models proprietary. Meta giving away comparable models for free is generous, but it’s also a business strategy that has limits.
The Open Source Tension
Here’s the uncomfortable truth that nobody in the AI industry wants to say out loud: pure open source AI is economically unsustainable at the frontier.
Training a frontier model costs hundreds of millions of dollars. If you release it for free, your competitors get the benefit of your investment without the cost. That works as a strategy when you’re trying to build an ecosystem (which Meta was). It stops working when the ecosystem is built and you need to monetize.
Meta’s likely approach: keep Llama open source as the “community” offering, while Avocado becomes the premium, proprietary model for enterprise customers. Think of it like Red Hat and Linux — the open source version is free, the enterprise version costs money.
This isn’t necessarily bad. It might actually be the only sustainable model for frontier AI development. But it does mean the era of “Meta gives away its best AI for free” is probably ending.
What This Means for Developers
If you’ve built on Llama, don’t panic. Meta isn’t going to pull the rug out. Llama will continue to be developed and released as open source. The community is too large and too valuable to abandon.
But here’s what you should be thinking about:
Diversify your model dependencies. If your entire stack depends on Llama, you’re exposed to Meta’s strategic decisions. Have a fallback plan with Mistral, Qwen, or other open models.
Watch the licensing. Llama’s license has always been “open” with asterisks (usage restrictions for large companies, no using it to train competing models). Future releases might have more restrictions.
The real competition is in fine-tuning. As base models commoditize, the value shifts to domain-specific fine-tuning and deployment optimization. That’s where you should be investing your effort.
The Bigger Picture
Meta’s AI strategy reflects a broader industry tension: everyone wants the benefits of open source (ecosystem, talent, goodwill) without the costs (giving away your competitive advantage).
Google open-sourced Gemma but keeps Gemini proprietary. Mistral started fully open but is increasingly offering proprietary enterprise models. Even Stability AI, which built its brand on open source, has struggled financially.
The pattern is clear: open source is a great growth strategy but a difficult business model. The companies that figure out how to do both — maintain a vibrant open ecosystem while building proprietary products on top — will win.
Meta is trying to figure this out in real time. The Avocado project is their first serious attempt at having it both ways. Whether it works will tell us a lot about the future of open source AI.
My bet: Meta will keep Llama competitive enough to maintain the ecosystem, while Avocado targets the enterprise market where companies are willing to pay for better performance, support, and SLAs. It’s not as idealistic as “open source everything,” but it might be the only approach that actually works long-term.
🕒 Last updated: · Originally published: March 12, 2026