Meta on Wednesday announced its first major model since CEO Mark Zuckerberg rebooted the company’s AI efforts last year under a new division called Meta Intelligence Labs. The model, called Muse Spark, is a step toward Zuckerberg’s vision of “personal superintelligence,” the company says, and for now, it will remain closed source.
Zuckerberg said in a social media post that Meta’s goal is to build AI products that “don’t just answer your questions but act as agents that do things for you”. The billionaire added that he is “optimistic that this will support a wave of creativity, entrepreneurship, growth, and health.”
Muse Spark certainly appears to be a major upgrade over Meta’s last big release, Llama 4, which came out in April 2025 and was viewed in the tech industry as a disappointment with middling performance.
Meta is making Muse Spark available via meta.ai and through the Meta AI app. Unlike Llama, Muse Spark is not being released for others to download, though the company says it hopes to open source future versions. Meta was previously seen as a leader in open source AI and made its Llama models available for researchers, startups, and hobbyists to download and customize.
“Looking ahead, we plan to release increasingly advanced models that push the frontier of intelligence and capabilities, including new open source models,” Zuckerberg wrote.
Meta’s self-reported benchmark scores for Muse Spark suggest the model is better at some tasks than the latest models from OpenAI, Anthropic, Google, and xAI. “Muse Spark is the first step on our scaling ladder,” Meta said in a blog post, referring to its goal of building AI that far outstrips human abilities.
Artificial Analysis, an AI benchmarking company that got early access to Muse Spark, said on social media that the new model is one of the best it has tested. “Muse Spark scores 52 on the Artificial Analysis Intelligence Index, placing it within the top 5 models we have benchmarked,” the company said in its post, citing its own rubric for scoring models that combines various third-party benchmarks.
Meta says the new model is natively multimodal, meaning that it has been trained to handle images, audio, and video as well as text. Muse Spark also features advanced reasoning capabilities, a key feature of the best AI models available today, and it was built from scratch to have strong coding capabilities. Meta described these features as the foundation for building ever-more capable models using modern machine learning methods.
Meta says that it built Muse Spark to be especially good at providing medical advice. “To improve Muse Spark’s health reasoning capabilities, we collaborated with over 1,000 physicians to curate training data that enables more factual and comprehensive responses,” the company said in its blog post.
Zuckerberg has spent a small fortune overhauling Meta’s artificial intelligence efforts since Llama 4 came out. The tech giant poached top AI engineers from competing firms with compensation packages worth hundreds of millions. It also spent billions to acquire or make major investments in a number of AI startups. Meta recruited Alexandr Wang, the CEO of Scale, an AI training company, to lead its AI efforts after investing $14.3 billion in the company.
Meta also published a document outlining its vision for safely scaling AI models to superhuman levels of performance. The company’s Advanced AI Scaling Framework outlines safety checks that the company will perform as its models become increasingly advanced.