Meta is quietly plotting a shift that would move it out of the “open-first” corner of the AI world and much closer to the commercial playbooks used by OpenAI and Google: people familiar with the plans say Meta is building a next-generation model, code-named Avocado, that it may keep closed and charge developers to use. If that happens, Meta’s flagship AI would stop being primarily an open community resource and start looking a lot more like a profit center engineered to satisfy investors as well as researchers.

That turnabout would be a major change for a company that spent much of the last few years throwing weight behind open models. Meta’s Llama series—culminating in Llama 4 earlier this year—was explicitly positioned as the alternative to walled-garden APIs: models and weights you could run yourself, fork, and build on. That openness won Meta goodwill with startups and academics who’d seen other big players seal their best systems behind metered endpoints. But the goodwill has limits when the models in question don’t land as hoped.

The moment that appears to have cracked the open-source narrative was Llama 4’s stumble and the messy aftermath for its larger sibling, the so-called “Behemoth.” Internal skepticism about whether Behemoth improved enough over earlier Llamas reportedly led to delays and, in some accounts, shelving—an embarrassment that seems to have prompted fresh leadership scrutiny and a rethink of strategy inside Meta’s AI ranks.

Avocado is the clearest expression of that rethink. According to reporting, the project lives inside a tightly controlled unit known as TBD Lab—people describe a team physically clustered close to Mark Zuckerberg’s office and stacked with expensive talent—where the plan is to treat Avocado as a frontier-scale model whose weights don’t leave Meta’s infrastructure. That technical choice matters: it lets Meta meter and monetize access (APIs, cloud partnerships, enterprise features) instead of letting developers download and run heavyweight models wherever they want. Reports say the company could unveil Avocado as soon as next spring.

The business logic is obvious. Meta is pouring enormous sums into chips, data centers, and AI-optimized infrastructure, and a proprietary, high-margin model gives executives something concrete to point to when justifying those outlays to investors. At the same time, insiders say the company hasn’t entirely closed the door on openness: smaller or older Llama releases could remain available for experimentation, while Avocado and its successors live behind a paywall—classic “freemium” platform economics applied to foundation models.

Personnel moves have underscored the shift. Meta’s recent hires and organizational reshuffle—most notably the arrival of Alexandr Wang from Scale AI and the creation of a Meta Superintelligence Labs umbrella—signal a desire to bring commercial discipline and closed-system experience into the heart of model development. Wang and other star recruits are reported to be on multiyear packages that make clear the company expects big, commercial returns on its AI bets.

For developers and academics who built business models or research agendas around Meta’s earlier, more permissive posture, Avocado will feel like a reminder of an old truth about platforms: openness is a wonderful way to build mindshare until the money on the table gets big enough to change incentives. That doesn’t mean Meta is abandoning collaboration or transparency entirely—the company still benefits politically and technically from open releases—but the center of gravity appears to be shifting toward proprietary, revenue-driven systems.

There are broader stakes here beyond Meta’s balance sheet. A move like this tightens the same competitive dynamic that pushed other big AI labs to monetize access: when the best models sit behind controlled endpoints, the costs of doing cutting-edge work rise for startups, and academic researchers face new barriers to replication. Regulators and policymakers watching concentration in the AI stack are likely to pay attention if the industry’s last major open-source champion starts to charge at the frontier.

What to watch next: whether Meta actually keeps Avocado’s weights on-premises or chooses a hybrid route; how the company prices access (developer tiers, enterprise packages, cloud partnerships); and the developer reaction—will startups adapt, pay up, or double down on alternative open stacks? Meta’s pivot is still a story in motion, but the direction is clear enough to be worth paying attention to: the company that once waved the banner for open models looks ready to build a flagship product that, for better or worse, is meant to make money.