Meta just scrapped its old AI playbook and launched something entirely new

Meta just scrapped its old AI playbook and launched something entirely new

Quick Reads

• Meta launched a brand-new AI model on April 8 called Muse Spark, its first release from an entirely new internal AI division called Meta Superintelligence Labs.
• This is a deliberate break from Llama, the open-source AI series Meta has built its AI identity on for years. Muse Spark is proprietary and closed.
• The model was built under Alexandr Wang, former CEO of data company Scale AI, whom Zuckerberg brought in after growing frustrated with how far Meta’s AI had fallen behind OpenAI and Anthropic.
• Muse Spark is free to use for now, but requires signing in with a Facebook or Instagram account, raising immediate questions about what Meta does with that data.
• A more powerful “Contemplating” mode is coming, designed to tackle harder problems by running multiple AI agents simultaneously.

Meta has spent years building Llama, its family of open-source AI models, as the foundation of its AI strategy. On April 8, the company quietly walked away from that identity. Mark Zuckerberg’s company released a new AI model called Muse Spark the first to come out of Meta Superintelligence Labs, a division created specifically because Zuckerberg was reportedly unhappy with the progress of Meta’s existing AI efforts and how they lagged behind OpenAI’s ChatGPT and Anthropic’s Claude. For Meta, this is not an update. It is a reset.

To lead Meta Superintelligence Labs, Meta recruited Alexandr Wang, the former co-founder and CEO of Scale AI, a company that specialises in preparing and labelling data for AI training. Alongside the hire, Meta invested $14.3 billion in Scale AI for a 49% stake in the company.

Meta is also positioning Muse Spark as a tool that can help users with health questions, an area that competitors, including Google and OpenAI, are also moving into. This is a strategically crowded space and one where errors carry real consequences.

The more significant feature is still coming. Meta plans to roll out a “Contemplating” mode that will allow the model to tackle more complex problems by using multiple AI agents working in parallel on the same task simultaneously, which the company says will generate faster results without drastically increasing how long users have to wait.

To use Muse Spark, users must log in with an existing Meta account, Facebook or Instagram. Meta does not explicitly say that personal information from those accounts will be used by the AI, but Meta’s track record makes that assurance worth scrutinising. The company generally trains on public user data, and Muse Spark is being positioned as a personal AI product, one designed to know you. That combination should prompt legitimate questions from anyone handing over access.

Meta’s competitors have historically placed their more capable AI models behind a paywall. It is currently unclear whether Meta will follow the same strategy with Muse Spark. For now, free access gives it a real shot at scale, and scale is exactly what Meta does well. If Muse Spark reaches even a fraction of Facebook and Instagram’s combined user base, the usage numbers would be significant regardless of how it benchmarks against rivals.

What this launch really signals, though, is that the previous version of Meta’s AI strategy has been quietly buried. Llama built Meta a reputation in the open-source AI world. Muse Spark is chasing something different: a proprietary model that competes directly with the products people already use every day from OpenAI and Anthropic. Zuckerberg put it plainly on Threads

Whether Muse Spark can actually deliver that or whether it will follow the pattern of an earlier Meta AI model launch that was quietly postponed in March 2026 after it fell short of rival performance benchmarks will become clear over the coming months.