Meta Debuts ‘Muse Spark’ AI, Challenging Rival Models From OpenAI, Google

After a bruising year that saw its Llama 4 models draw yawns from developers and accusations of benchmark gaming, the social media giant on Wednesday unveiled Muse Spark.

The model is the first to emerge from the company’s newly formed Superintelligence Labs, led by Alexandr Wang. It follows months of internal restructuring and billions in investment aimed at catching up with rivals like OpenAI, Anthropic, and Google.

Meta described Muse Spark as “the first step on our scaling ladder” toward what it calls personal superintelligence.

Originally code-named “Avocado,” Muse Spark is built to be a “natively multimodal” powerhouse. This means it doesn’t just stitch images and text together; it processes voice, text, and visuals simultaneously. It is already live on the Meta AI web portal and mobile app, with plans to integrate it into WhatsApp, Instagram, and even Meta’s smart glasses in the coming weeks.

What Muse Spark can do

Meta says the model focuses on reasoning and real-world usefulness, particularly in areas like science, math, and health. It can also support interactive experiences, such as troubleshooting devices or generating simple games.

One standout feature is its ability to seamlessly combine different types of data. As Meta explains, the model is built “from the ground up to integrate visual information across domains and tools.”

The company also highlights health-related use cases. To improve accuracy, Meta says it worked with over 1,000 physicians to train the system, enabling it to explain topics like nutrition and exercise in more detail.

A Shopping mode is also in the works, pulling from creator content across Instagram and Threads to offer personalized product recommendations, a feature that leans directly into Meta’s advertising DNA.

New ‘contemplating mode’ for hard problems

A key addition is something Meta calls “Contemplating mode.” This feature uses multiple AI agents working in parallel to solve complex problems.

Meta says this approach helps the model compete with advanced reasoning systems from rivals. In internal testing, the mode achieved 58% on Humanity’s Last Exam and 38% on FrontierScience Research, according to the company’s technical post.

Explaining the concept, Meta noted, “To spend more test-time reasoning without drastically increasing latency, we can scale the number of parallel agents that collaborate to solve hard problems.”

How it stacks up

Meta isn’t claiming Muse Spark has claimed the AI crown. The company acknowledges gaps remain in coding and what it calls “long-horizon agentic systems.”

But independent testing by Artificial Analysis shows that Meta has considerably closed the gap with frontier competitors. Muse Spark scored 52 on the Intelligence Index, placing it in the top five models tested globally, trailing only Gemini 3.1 Pro Preview, GPT-5.4, and Claude Opus 4.6. For context, Meta’s previous Llama 4 Maverick managed just 18 points.

Where Muse Spark particularly shines is efficiency. Artificial Analysis found the model burned through 58 million output tokens to complete its full Intelligence Index run, on par with Gemini 3.1 Pro Preview and dramatically less than Claude Opus 4.6 (157 million) or GPT-5.4 (120 million).

The company attributes this to a phenomenon it calls “thought compression.” In its blog, Meta explained: “After an initial period where the model improves by thinking longer, the length penalty causes thought compression — Muse Spark compresses its reasoning to solve problems using significantly fewer tokens.”

Goodbye, Open Source?

For years, Meta was the poster child for “open-source” AI, letting developers download and run their Llama models for free. Muse Spark changes that. For now, the model is proprietary, meaning the code stays behind Meta’s walls.

While Wang mentioned on X that there are “plans to open-source future versions,” the current focus is on a private API preview for select partners and potential future subscriptions.

For a deeper look at how AI is reshaping cybersecurity, see our coverage of Anthropic’s Project Glasswing, a tightly controlled effort to hunt down critical software vulnerabilities before attackers do.