Meta launches Muse Spark AI model for multimodal reasoning and orchestration

Meta has released its first major Artificial Intelligence (AI) model, Muse Spark, developed by Meta Superintelligence Labs, led by Alexandr Wang. It is a multimodal reasoning model, meaning it can work with different types of data, including text and images.

"Muse Spark is the first step on our scaling ladder and the first product of a ground-up overhaul of our AI efforts," Meta stated in a blog post.

Muse Spark is designed to think through problems step by step. It can also use tools, present its reasoning visually, and coordinate multiple AI systems simultaneously. Here’s everything you need to know about Meta’s new Muse Spark AI model.

Also read: Anthropic releases preview of new AI model Mythos for cyber defence

Muse Spark: Features and performance

Muse Spark is the first “inaugural model” launched under Meta’s Superintelligence Labs, which was formed by CEO Mark Zuckerberg in July 2025. This new model is said to offer “personal intelligence” for everyday use, designed to manage tasks such as visual understanding, health, shopping, and social content.

Zuckerberg, in a Facebook post, revealed that Meta is “ building products that don't just answer your questions but act as agents that do things for you.” The new AI model is rolling out in web and the Meta AI app. The company states that the model is built to improve over time, learning from interactions and becoming more capable as it is used.

The AI model also comes with a “Contemplating mode,” where the AI uses multiple agents at the same time to analyse a problem. These agents work in parallel, which helps the model reason more efficiently. Therefore, this makes the model more powerful at solving complex problems. The model is said to achieve 58.4% in Humanity’s Last Exam and 38.3% in Frontier Science Research.

The Musk Spark also comes with health reasoning capabilities, for which the company have collaborated with over 1,000 physicians for training data. This feature canuser provide health information such as the nutritional content of foods or muscles activated during exercise. 

In terms of benchmark, Musk Spark has achieved 42.1% on HealthBench Hard, which is slightly higher than GPT 5.4, but way ahead of other models like Gemini 3.1 Pro. On DeepSearchQA, the model scored 74.8%, placing it ahead of some competitors but slightly behind the top performers. In SWE-Bench Verified, it achieves 77.4%, close to leading scores in the low 80s range. For the more challenging SWE-Bench Pro, it records 52.4, again trailing slightly behind the best score of 57.7.

“Looking ahead, we plan to release increasingly advanced models that push the frontier of intelligence and capabilities, including new open source models,” Zuckerberg said. “We are building products that don’t just answer your questions but act as agents that do things for you.”