Meta just broke its open source AI promise. Yesterday, the company announced Muse Spark — its first proprietary AI model with closed weights and API-only access. This is the same company whose CEO Mark Zuckerberg wrote “Open Source AI is the Path Forward” just 18 months ago in July 2024. What changed? The $14.3 billion answer matters to every developer betting on the future of accessible AI.
From Open Source Champion to Proprietary Player in 18 Months
The timeline tells the story. In July 2024, Zuckerberg published a manifesto arguing that “open source AI represents the world’s best shot at harnessing this technology to create the greatest economic opportunity and security for everyone.” He compared the trajectory to Unix versus Linux, predicting that open source would eventually dominate just as Linux did in the server market.
Fast forward to June 2025: Meta spends $14.3 billion to acquire a 49% stake in Scale AI and brings in its co-founder Alexandr Wang as Meta’s first Chief AI Officer. Wang, who built Scale AI’s business serving closed-model companies like OpenAI and Anthropic, takes over a new division called Superintelligence Labs with zero open source background.
Nine months later, we have Muse Spark — Wang’s first major model, and it’s completely closed. No weights, invitation-only API access, undisclosed pricing. Meta now calls this a “hybrid strategy” where they’ll release both open and proprietary models. Translation: the cutting-edge stuff stays locked down.
What Developers Lost When the Weights Closed
The difference between open and closed weights isn’t academic. When Llama 3.1 had open weights, developers could self-host without API costs, fine-tune for custom use cases, audit model architecture, and verify benchmark claims independently. None of that is possible with Muse Spark.
Now developers face vendor lock-in. Meta controls access through invitation-only APIs with unknown pricing. You can’t customize the model, can’t run it locally, and can’t verify if it actually performs as claimed. That last point matters — Meta faced criticism over its Llama 4 benchmarking practices. With closed weights, you just have to trust them.
The irony: by restricting access, Meta controls the narrative without independent verification. The very transparency Zuckerberg championed in 2024 is now optional.
The End of Frontier Open Source AI?
Meta was the last major tech company releasing truly open weights at frontier scale. OpenAI, Anthropic, and Google have always kept their best models proprietary. Meta’s Llama series was the exception — the one place developers could access cutting-edge AI without gatekeepers or API bills.
Muse Spark changes the equation. If Meta’s “hybrid strategy” means Muse gets the frontier capabilities while Llama becomes the “good enough” open option, we’ve lost something important. The message to the market is clear: you can’t compete at the frontier without going closed.
Zuckerberg’s Unix/Linux analogy from 2024 feels hollow now. He argued that open source would eventually overtake proprietary AI just as Linux overtook Unix. However, in 2026, Meta itself abandoned the open source path for its most ambitious AI work. If the company that spent billions championing open source can’t make it work competitively, what does that say about the model?
What Comes Next
The Muse Spark announcement frames this as “the first step on our scaling ladder toward personal superintelligence.” That phrasing suggests more Muse models are coming, likely all closed. Llama may continue to exist, but as a second-tier option for developers who can’t access or afford the proprietary tier.
The AI industry is worse off for it. Fewer open options mean less innovation, higher barriers to entry, and more corporate control over who gets to build with frontier AI. Meta built its credibility by releasing Llama openly — was that just marketing to attract AI talent and training data before pivoting to a proprietary business model?
We’ll find out. But this moment — when the last major open source frontier AI provider joined the proprietary club — might be one we regret.
