Meta Drops Its Open-Source Playbook and Launches a Closed AI Model Called Muse Spark

Meta Drops Its Open-Source Playbook and Launches a Closed AI Model Called Muse Spark

Quick Reads

• Meta has launched Muse Spark, its first proprietary AI model, marking a sharp departure from its previous open-source approach with the Llama family.
• The model is the first product out of Meta Superintelligence Labs, a unit created after CEO Mark Zuckerberg grew frustrated with Llama’s performance against rivals like OpenAI and Google.
•  Alexandr Wang, the 29-year-old former Scale AI co-founder hired nine months ago for $14.3 billion for a 49% stake in his data company, leads the lab that built Muse Spark.
• Unlike Llama, which anyone could freely download and run, Muse Spark is locked behind Meta’s own app and website, with API access offered only by invitation to select developers.
• Meta says the model rivals top AI systems from OpenAI, Anthropic and Google, though the company has previously been criticised for inflating benchmark results on past models.

For three years, Meta’s identity in the AI race was straightforward: give developers free access to powerful models they could download, modify, and run on their own computers. That approach, delivered through the Llama series of models, earned Meta enormous goodwill in the tech community and positioned it as the champion of open-source AI. By last year, self-hosting Llama models offered businesses an 88% cost reduction compared to using paid API providers from rivals.

But that era appears to be over, at least for now.

Muse Spark is proprietary, with the company saying only that it hopes to open-source future versions of the model. For now, access is limited to the Meta AI app and website, and developers who want to build with it must apply for an invitation to a private API preview.

The pivot follows a bruising period for Meta’s AI ambitions. Llama 4 debuted to mixed reviews last year, with Meta ultimately admitting to gaming benchmarks a serious credibility problem in an industry that relies on those tests to compare models. That stumble reportedly pushed Zuckerberg to act. He recruited Alexandr Wang, the former co-founder and CEO of Scale AI, to lead Meta Superintelligence Labs, and invested $14.3 billion in Wang’s data labelling company for a 49% stake.

Over the past nine months, Meta’s team rebuilt its AI stack from the ground up, including improvements to model architecture, optimisation, and data curation.

The result, according to Meta, is a model that can see and understand images, reason through complex problems, and run multiple AI agents simultaneously to tackle difficult tasks faster. The company plans to roll out a “Contemplating” mode that uses multiple AI agents working on the same problem in parallel to generate faster results on harder queries. Think of it as giving the AI multiple brains to consult at once, rather than thinking through a problem alone in sequence.

Muse Spark will debut inside Facebook, Instagram, WhatsApp, Messenger, and Meta’s Ray-Ban AI smart glasses in the coming weeks.

On the efficiency front, Meta reports that Muse Spark achieves its reasoning capabilities using over an order of magnitude less compute than Llama 4 Maverick, driven by a process called “thought compression” where the model is penalised for excessive reasoning time during training, forcing it to solve problems with fewer steps without sacrificing accuracy.