Meta Launches Muse Spark, Its First Superintelligence Model

Key insights

• Meta spent hundreds of billions on AI infrastructure but deliberately launched a limited first model, signaling a strategic shift away from overpromising after the Llama 4 Maverick disappointment.
• Plans for a paid API would mark an entirely new business model for Meta, moving it from free open-source provider to direct competitor of Anthropic and OpenAI.
• Benchmarks for the unreleased Contemplating Mode (58% on Humanity's Last Exam) signal frontier-level capability to investors before the model ships, a calculated confidence move.
• Muse Spark is purpose-built for Meta's own apps first: shopping, creators, and coding. This makes it an ecosystem play as much as a general AI product.

In Brief

Meta has launched Muse Spark, the first AI model from Meta Superintelligence Labs (MSL), the research group that Chief Executive Officer (CEO) Mark Zuckerberg built around Chief AI Officer Alexandr Wang. The model is live today in Meta's AI app and website, with a wider rollout to Instagram, WhatsApp, and Meta's AI glasses coming in the weeks ahead. CNBC's Julia Boorstin broke the news on air as Meta's stock shot up 6.5%.

A Cautious First Step — by Design

If you were expecting Meta to come out swinging with a ChatGPT-killer, Muse Spark is not quite that. Meta calls it their most powerful model yet, but the company is deliberately framing this as "the first step on our scaling ladder" rather than a finished product. That careful framing is itself a strategy.

After the Llama 4 Maverick release earlier this year drew criticism for underdelivering on its benchmarks, Meta appears to be learning from that experience. The approach here: ship something real, keep expectations modest, and let the product speak for itself.

The model was originally codenamed Avocado inside Meta, which gives you a sense of how it was treated internally: a quiet, iterative project rather than a splashy announcement. The public name, Muse Spark, is part of a broader Muse family of models Meta says it is building out over time.

What Muse Spark Can Actually Do

Muse Spark is a multimodal model, meaning it can work with text, images, and code at the same time rather than handling them separately. This is not an add-on: Meta says it was built from the ground up to handle different types of input together.

The initial focus is on features that make sense inside Meta's existing apps:

• Shopping mode surfaces product ideas from creators a user already follows, turning the social feed into something closer to a personalized storefront.
• Visual coding tools let users describe a website or mini-game in plain language and watch the model build it.
• Creator discovery helps users find relevant content from the accounts they already care about.

Beyond these consumer-facing features, the model includes several capabilities drawn from Meta's technical blog post. Visual chain of thought lets the model reason about images step by step, the way you might annotate a diagram before drawing a conclusion. Tool-use lets it call external services and APIs mid-conversation. Health reasoning is trained on data from over 1,000 physicians, so it can give factual, grounded answers to health questions rather than vague disclaimers.

On the engineering side, Meta claims an order-of-magnitude improvement in compute efficiency compared to Llama 4 Maverick. That means Muse Spark can do more with less processing power, partly through a technique called thought compression, where the model trims down the number of "thinking steps" it takes internally without sacrificing accuracy.

The Bigger Model Is Coming Later

The most intriguing part of the announcement is what Meta chose not to release. Alongside Muse Spark, the company teased a more advanced mode called Contemplating Mode, a feature where the AI takes extra time to reason through hard problems before responding, similar to how OpenAI's "o" series models work.

Meta says Contemplating Mode is already competitive with frontier models like Google's Gemini, DeepSeek's DeepThink, and OpenAI's GPT o3 Pro. It scores 58% on a benchmark called Humanity's Last Exam, a notoriously difficult test of reasoning across science, math, and logic, and 38% on FrontierScience Research, another hard benchmark used to compare cutting-edge models. These are frontier-level numbers.

But the mode is not available yet. Meta is releasing the benchmarks now and shipping the feature later, a move that functions as a signal to investors and a warning to competitors: the next release is going to be significantly more capable.

A New Revenue Stream for Meta

Alexandr Wang spent nine months building Meta Superintelligence Labs, backed by hundreds of billions of dollars in AI infrastructure investment. That infrastructure includes the Hyperion data center, which Meta describes as a key part of its AI stack from model training all the way to deployment.

Now, for the first time, Meta is talking about turning that investment into a direct revenue line. The company says it plans to open paid API access to Muse Spark, similar to how Anthropic charges developers to use Claude and Google charges for Gemini access. An API (Application Programming Interface) is essentially a way for other software products to tap into the model's capabilities, letting developers build on top of it rather than building their own.

This is new territory for Meta. The company has historically positioned its AI models, particularly the Llama series, as free and open-source alternatives to the paid competitors. Charging developers for API access would make Meta a direct commercial competitor to OpenAI and Anthropic, not just a free option in the ecosystem. A private API preview is already opening to select developers today.

Meta shares closed the day up 6.5%, reflecting investor enthusiasm for what one CNBC commentator described as a "transformation happening in real time."

What This Means

Muse Spark is not a ChatGPT rival yet. But the combination of a real product launch, ambitious benchmark numbers for an upcoming mode, and a clear pivot toward paid API revenue suggests Meta is building toward something more significant. The cautious rollout buys Wang time to get the bigger models right. And if the Contemplating Mode benchmarks hold up when it ships, Meta's position in the frontier AI race will look very different by the end of the year.

Glossary

Sources and resources