What happens when a smaller, speed-driven AI team inside a tech titan decides the old playbook won’t cut it anymore? At Meta, that question has stopped being academic and started shaping product strategy—and tension.
A new model, a new fault line
Reports from multiple outlets this week say Meta is quietly building an internal model called “Avocado,” and that it may not be open source. The work is happening inside a compact unit known as TBD inside Meta’s AI Superintelligence Labs, led by Chief AI Officer Alexandr Wang. For years Meta loudly championed open-source release of models like Llama; now parts of the company appear to be steering toward closed, proprietary systems.
Those reports also describe cultural friction: an “us-versus-them” divide between the newer AI hires clustered around Wang and longstanding Meta lieutenants who long stewarded the company’s engineering and product culture. The New York Times, among others, described clashes with veteran leaders—people who built Meta’s infrastructure and product orgs—over how aggressive and how closed the company’s next-generation AI efforts should be.
Why this pivot matters
Open-sourcing models did a lot for Meta: it built community goodwill, accelerated research, and turbocharged third-party innovation around Llama. But it also handed rivals and bad actors access to powerful capabilities. Internally, engineers who favor closed models argue commercialization and safety require tighter control. Externally, the competitive landscape—OpenAI, Google, Anthropic—has pushed big players to prioritize defensible products and revenue streams over broad releases.
The tug-of-war played out in product delays too. Meta’s Llama 4 “Behemoth” release was reportedly pushed back and debated internally. Some execs considered abandoning it, as developers and partners voiced frustration with what was finally available. Those technical and market pressures make Avocado less an academic exercise and more a strategic pivot.
People and politics
This isn’t just a debate about code licensing. It’s about organizational identity. Yann LeCun, a high-profile Meta figure and open-source advocate, recently left the company, and hundreds were laid off from FAIR earlier this year—moves that signaled a broader reshuffle of priorities. Veteran leaders such as Andrew Bosworth (and others tied to Meta’s product and hardware road maps) reportedly have clashed over how closed or open the company should be.
Money talks. Meta’s AI ambitions carry enormous price tags, and executives are under pressure to show returns. Investors responded: the stock nudged lower after initial reports about a shift toward a closed model, a modest market reaction that nonetheless underlines how high the stakes are for a company that has been courting both developer communities and enterprise customers.
The calculus: safety, speed, and sales
There are three practical drivers behind the move:
- Safety and control: Closed models make it easier to monitor misuse, enforce guardrails, and restrict access to potentially dangerous capabilities.
- Commercialization: Proprietary models are simpler to monetize—licensing, APIs, product bundles—than widely distributed open models.
- Competitive pressure: Keeping pace with rivals that produce tightly integrated, revenue-generating AI services incentivizes a pivot away from pure openness.
Meta isn’t alone in balancing these forces. The industry is experimenting with different trade-offs between openness and productization. For instance, Microsoft has been pushing its own specialized image model—another sign of companies sharpening distinct offerings rather than sharing everything openly Microsoft’s MAI-Image-1. And as companies make AI more agentic and product-centric, features that book appointments or act on a user’s behalf are pushing firms to lock down integrations and user data flows, not loosen them Google’s AI Mode agentic booking.
What developers, partners and users should watch
If Meta pursues Avocado as a closed model, the developer ecosystem around Llama will likely recalibrate: some tooling and startups may double down on open alternatives, others will chase paid integrations or ports. Enterprises that value control and service-level guarantees might welcome a proprietary Avocado offering, while researchers and hobbyists could lose a valuable sandbox.
Expect three near-term signals to watch for: public licensing changes for Llama, how Meta prices and packages any Avocado API, and whether the company releases safety and transparency documentation that reassures regulators and partners. Meanwhile, the broader industry continues to roll out specialized capabilities—Google’s and Microsoft’s moves show the market is fragmenting into differentiated stacks rather than a single open frontier Gemini’s deeper integrations.
Meta’s internal story right now is one of reinvention and infighting in equal measure. Whether Avocado becomes a closed, revenue-driving crown jewel or another shelved experiment depends on technical results, board-level appetite, and if the company can stitch together a culture that tolerates both speed and deliberation. Either way, the debate inside Meta is a clear sign: open-source as a guiding doctrine for AI is no longer a settled question, and the industry is rewriting the rulebook in real time.