Meta’s Next AI Models Will Be Open Source. Is That Generosity or the Smartest Competitive Move of the Year?

For more than a year, a silent recalibration has been taking place somewhere on Meta’s Menlo Park campus, behind the glass walls and the well-fed engineers. The release of Llama 4 did not go as planned for Zuckerberg. Anyone who keeps a close eye on this industry could sense it. The company that once defined social media appeared to be running second, possibly third, in a race it had publicly committed billions of dollars to winning. The benchmarks were acceptable, but the response was muted.

The hiring frenzy followed. The keys to a new “superintelligence” team were given to Alex Wang, who was taken from Scale AI in a deal worth over fourteen billion dollars. Pay packages that resembled NBA contracts rather than tech salaries were allegedly offered to engineers. And now, following months of conjecture, that team’s first product has emerged: Muse Spark, the first act of an internally codenamed series called Avocado. Meta claims that it is small, quick, and able to reason through questions related to math, science, and health.

It’s not just the model that’s intriguing. It’s the tactic that surrounds it. In keeping with the tradition started with Llama, Meta has confirmed that larger versions of Muse Spark will be made publicly available. This appears to be generosity at first glance. Researchers and startups who could never afford to license OpenAI or Anthropic’s most sophisticated systems can access free models and open weights. But in this field, generosity hardly ever explains anything. Underneath, the logic is typically sharper.

Think about the economics. Despite costing roughly six times more than open alternatives, closed AI models account for about 80% of all tokens processed on platforms like OpenRouter, according to a recent paper co-authored by Frank Nagle at MIT. At release, open models perform about 90% better than their closed counterparts, and the difference usually narrows quickly. According to the paper, the global AI economy could save about $25 billion annually by reallocating demand to open models. It’s not a rounding error.

It’s difficult to ignore the pattern as you watch this develop. At least not yet, Meta cannot simply outperform OpenAI in terms of raw model performance. Muse Spark tied for fourth on the Artificial Analysis index, falling behind in coding and abstract reasoning but catching up in language and visual comprehension. Wang acknowledged that “rough edges” needed to be polished. Making the leader’s business model more difficult to defend is the next best option if you are unable to outbuild them.

In 2026, open source primarily accomplishes this. Developers are not paying OpenAI when they build on a free Meta model. One fewer client is renewing an Anthropic contract for each business that optimizes Avocado for its own workflows. Similar work is being done from a different perspective by NVIDIA’s Nemotron family. the compounds under pressure.

Additionally, Meta discusses the user data angle in a less transparent manner. With 3.5 billion users on Facebook, Instagram, and WhatsApp in addition to smart glasses becoming more and more commonplace, the company has a distribution advantage that is nearly unmatched by competitors. These aren’t moonshots; Meta AI has built-in shopping features that allow users to superimpose a mug onto their shelf or calculate the calories in a meal from a picture. They are hooks. quiet ones.

It remains to be seen if it is effective. Instead of being persuaded, investors appear cautiously curious. Earlier this year, Zuckerberg informed them that the initial models would be “good, but more importantly” would demonstrate the team’s progress. That’s how someone who manages expectations speaks. Although the parallel isn’t perfect, Tesla had to deal with similar doubts years ago. Open source may prove to be Meta’s most methodical decision in a long time. Or it could be the most costly method of giving something away ever devised.