Meta Muse Spark Proprietary AI Model Marks a New Era for Zuckerberg
Quick Reads
- Meta has launched Muse Spark, its first-ever proprietary AI model, ending years of open-source AI releases.
- The model is developed by Meta Superintelligence Labs, led by Chief AI Officer Alexandr Wang.
- Muse Spark is now available on the Meta AI website and app, with a private API preview for select developers.
- The model can see, reason, and handle health questions and is already powering Meta’s AI assistant.
- Meta says it plans to spend between $115 billion and $135 billion on AI infrastructure in 2026.
Meta has launched Muse Spark, its first-ever proprietary AI model, ending years of open-source AI releases. The model is developed by Meta Superintelligence Labs, led by Chief AI Officer Alexandr Wang. Muse Spark is now available on the Meta AI website and app, with a private API preview for select developers. The model can see, reason, and handle health questions and is already powering Meta’s AI assistant. Meta says it plans to spend between $115 billion and $135 billion on AI infrastructure in 2026.
From Free to Locked: What Muse Spark Actually Is
Muse Spark is a reasoning model. Think of it as an AI that does not just answer questions it thinks through problems step by step before responding. It can process text, images, and voice inputs all at once. That is what engineers mean when they call it “natively multimodal.”
The model is built by Meta Superintelligence Labs (MSL), a new division created in June 2025. Mark Zuckerberg set it up after Meta’s previous flagship, Llama 4, received a poor reception from developers. He recruited Alexandr Wang, then CEO of data company Scale AI, to lead it. Meta invested $14.3 billion in Scale AI for a 49% stake as part of that deal.
Wang and his team spent nine months rebuilding Meta’s entire AI stack from scratch. The result is Muse Spark, a model that, according to Meta, achieves similar capabilities to Llama 4 Maverick using over ten times less computing power.
That efficiency matters. Meta serves over three billion users across its apps. Running a powerful AI model at a fraction of the usual cost changes the economics of the entire operation.
What Makes This Launch Different for Developers
Until now, Meta was the loudest champion of open-source AI. Its Llama models could be downloaded freely and customised by anyone. By early 2026, the Llama series had reached 1.2 billion downloads. Developers used it to build their own AI products at a fraction of the cost of using paid rivals.
Muse Spark breaks that model entirely. No architecture details have been released. No weights are available for download. Access is through Meta’s apps or a private, invite-only API preview. When VentureBeat asked Meta directly whether Llama development had ended, a company spokesperson only confirmed that existing Llama models would remain available without saying a word about future releases.
Wang acknowledged the shift publicly, writing on X: “Nine months ago we rebuilt our AI stack from scratch. New infrastructure, new architecture, new data pipelines. This is step one. Bigger models are already in development with plans to open-source future versions.”
The developer community has responded with caution. For many, Llama was attractive because it could be fine-tuned for specific use cases. Muse Spark cannot be customised in the same way.
How Does Muse Spark Perform?
On independent benchmarks, Muse Spark scores 52 on the Artificial Analysis Intelligence Index. That places it fourth globally behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6. Meta has not claimed to have built the world’s best model.
Its strongest results are in health-related tasks. Meta worked with over 1,000 physicians to build that capability. The model scored 42.8% on HealthBench Hard, the highest result among frontier models in that category. It can explain nutritional data, describe medication effects, and display muscle activity during exercise.
It also features a “Contemplating mode,” which runs multiple AI agents in parallel to tackle complex problems. According to Meta’s own data, this mode scored 58% on Humanity’s Last Exam, a benchmark designed to measure near-expert-level reasoning.
Where it struggles is in coding. It scored 59.0 on Terminal-Bench 2.0, well below GPT-5.4’s 75.1. Meta has acknowledged this gap.
Meta stock rose more than 9% on the day of the launch. Investors read it as evidence that the $14.3 billion spent on Wang’s team had produced something real. But the business model is still unproven. OpenAI and Anthropic are collectively valued at over $1 trillion. Meta has yet to establish a major revenue stream from AI outside its advertising business.
Muse Spark is rolling out inside Facebook, Instagram, WhatsApp, Messenger, and Meta’s Ray-Ban AI glasses in the coming weeks. Meta also plans to use the model to power a shopping feature that connects users to products based on their interests across its platforms. The privacy implications of that are real and worth watching the model requires users to log in with a Meta account.
The global AI market is projected to grow from roughly $22 billion in 2025 to nearly $325 billion by 2033, according to Grand View Research. Meta is making a massive bet that Muse Spark is its entry point into that future.
