Meta Launches Muse Spark, Its First Model From the New Superintelligence Lab

What Muse Spark Is and Why It Matters

For years, Meta has been the company that shows up to the AI race a step behind. Its open-source Llama models earned developer goodwill but never quite matched the frontier models from Anthropic, Google, and OpenAI where it counted. Last week, Meta tried to change that narrative with the launch of Muse Spark, its new flagship AI model and the first to come out of the superintelligence lab Mark Zuckerberg spent billions building in 2025.

Muse Spark Closes the Gap on Reasoning, But Not on Coding

The model shows real progress. On writing and reasoning benchmarks, Muse Spark performed significantly better than Meta's previous models and came close to the top offerings from its biggest competitors. That's not nothing. Closing the gap on reasoning is exactly what Meta needed to do to stay credible in a space where the bar keeps moving.

But the gap that matters most right now is coding, and Muse Spark still lags there. That's a problem because coding ability has become the defining benchmark of the current AI moment. Anthropic in particular has made it a centerpiece of its Claude models, and the day before Meta's announcement, Anthropic revealed its latest model, Mythos, was considered too powerful to release publicly due to cybersecurity concerns. Whether or not that framing is marketing, it sets a tone that Meta is still working to match.

A Closed Model From the Company That Built Its Rep on Open Source

Muse Spark, which was known internally as Avocado during development, is available now on Meta's standalone AI app and will roll out to WhatsApp, Instagram, and Meta's AI smart glasses in the coming weeks. One notable departure from Meta's usual playbook: the model is closed source. Meta has built much of its developer reputation on open-sourcing its models, so keeping Muse Spark proprietary is a meaningful shift, even if the company left the door open to open-sourcing parts of it later.

The Investment Behind the Model

The stakes here are higher than a single model launch. Zuckerberg has pledged $600 billion in data center investment and projected spending of up to $135 billion this year alone, nearly double what Meta spent in 2024. He brought in Alexandr Wang, the 29-year-old Scale AI founder, as chief AI officer to lead the push. Muse Spark is the first public proof of concept for all of that. As Forrester research director Mike Proulx put it, the model is "really at the center of Meta's AI credibility."

What Comes Next

Internally, the development was reportedly a nine-month sprint marked by delays and tension. Zuckerberg himself walked back expectations in January, saying the model would demonstrate trajectory rather than push the frontier. Meta echoed that publicly at launch, framing Muse Spark as a fast, capable foundation rather than a ceiling. Its next model, codenamed Watermelon, is already in development.

That framing is probably the right one. Muse Spark is not the model that puts Meta on top of the AI stack. It is the model that proves Meta's new AI organization can actually ship something competitive. Whether Watermelon is the one that changes the conversation is the more interesting question, and the answer will tell us a lot about whether Zuckerberg's AI spending spree is paying off or just padding the infrastructure bill.