
Meta Built Its Entire AI Brand on Open Source. Then It Launched Muse Spark and Locked the Door.
Meta just abandoned its open source identity to launch a proprietary AI model. The company that preached openness built a walled garden.
The AI Post newsroom — delivering AI news at the speed of intelligence.
For the better part of three years, Meta had exactly one selling point in the AI race: we give it away for free. While OpenAI charged $200 a month and Anthropic locked its best models behind enterprise contracts, Meta released Llama under open licenses and let the entire developer world build on top of it. It was the most effective branding play in AI. Meta was the people’s lab.
That era ended on Wednesday.
Meta officially unveiled Muse Spark, the first model from its $15 billion Superintelligence Labs unit led by former Scale AI CEO Alexandr Wang. It is proprietary. Not open weight. Not open source. Not downloadable. If you want to use it, you log into Meta AI with your Facebook or Instagram account, and you use it inside Meta’s ecosystem. A few "select partners" get API access in a private preview. Everyone else waits.
Meta’s blog post includes the most corporate sentence imaginable: "We hope to open-source future versions of the model." Hope. Future. Versions. That is three hedges in eight words.
Here is what actually happened. Llama 4 flopped. Badly. The benchmarks were manipulated, developers noticed, and Meta lost the credibility it had built with the open source community almost overnight. Zuckerberg responded by spending $14.3 billion on Scale AI, poaching researchers from OpenAI and Anthropic, and creating a completely new AI division. Nine months and hundreds of billions in infrastructure commitments later, the result is a model that Meta itself describes as "small and fast by design" and acknowledges has "performance gaps" in agentic systems and coding.
To be fair, Muse Spark is not bad. The benchmarks show it is competitive across multimodal reasoning, science, math, and health queries. It scored 89.5% on GPQA Diamond, trailing Gemini 3.1 Pro (94.3%), Claude Opus 4.6 (92.7%), and GPT-5.4 (92.8%). On HealthBench Hard, it beat all of them. The "Contemplating" mode, which spins up multiple AI agents to reason in parallel, is genuinely interesting.
But none of that changes what this really is: a confession.
Meta went proprietary because it realized giving away the AI model was giving away the moat. When Google released Gemma 4 under Apache 2.0 and it immediately started eating Llama’s developer share, Meta saw the problem clearly: if anyone can replicate your open model, your open model is not a competitive advantage. It is a training dataset for your competitors.
The timing tells the real story. OpenAI and Anthropic are now collectively valued at over $1 trillion. Google’s Gemini has gained serious traction. Meta’s stock popped 6.5% on the Muse Spark announcement, but that was also the day Trump suspended Iran attacks and the entire market rallied. Meta is not leading the AI race. It is spending $115 to $135 billion in capex this year just to stay in it.
The most revealing detail in the CNBC report: Meta is experimenting with selling API access to Muse Spark as a new revenue stream. Meta. The company that ridiculed paid AI models as gatekeeping. Now charging for API access. You cannot write satire this good.
There is also a privacy dimension nobody is talking about. Muse Spark requires a Meta account to use. That means your Facebook or Instagram data is one login screen away from your AI conversations. Meta does not explicitly say it will use personal data from those accounts, but this is Meta. The company that got caught training AI on public user data. The idea that there is a firewall between your Instagram history and your Muse Spark health queries requires a level of trust that Meta has not earned.
Zuckerberg wrote on Threads: "We are building products that don’t just answer your questions but act as agents that do things for you." Translation: we need your data inside our ecosystem to build the agents that will eventually replace the apps you use to avoid giving us your data.
Look, I understand the business logic. Open source was not making Meta money. Llama was a loss leader that stopped leading. The AI model market is winner-take-most, and Meta cannot afford to keep giving away the thing everyone else charges for. But let us not pretend this is anything other than what it is: the biggest open source U-turn in tech history.
The developer community that built on Llama, that evangelized Meta’s approach, that chose Meta specifically because it was open? They just got the Facebook treatment. Come for the free platform. Stay because your data is locked in.
First reported by CNBC, Reuters, and TechCrunch.