On Wednesday, Meta unveiled its first flagship large artificial intelligence model since CEO Mark Zuckerberg restructured the company’s AI work last year, grouping all related initiatives under a new dedicated division called Meta Intelligence Labs. Dubbed Muse Spark, the new system marks a key milestone toward Zuckerberg’s long-term vision of building “personal superintelligence,” the company confirmed, and will remain closed-source for the time being.
In a public social media post, Zuckerberg outlined that Meta’s core goal is to build AI tools that “do more than just answer your questions—they act as autonomous agents that complete tasks on your behalf.” The Meta CEO added he is optimistic this new generation of AI will unlock a widespread wave of growth across creative work, entrepreneurship, economic expansion, and public health.
Industry observers broadly view Muse Spark as a major upgrade over Meta’s previous flagship release, Llama 4, which launched in April 2025. That 2025 model was widely considered a disappointment across the tech sector, with underwhelming, middling performance across key benchmarks.
Muse Spark is currently accessible to end users through Meta’s meta.ai website and the standalone Meta AI mobile app. Unlike the Llama model line, Muse Spark is not available for external developers to download and run on their own infrastructure, though Meta notes it hopes to release open-source versions of future iterations of the model. For years, Meta has been recognized as a leader in open-source AI development, having previously made all generations of its Llama models available for download and customization by researchers, startups, and independent AI hobbyists.
“Looking ahead, we plan to release increasingly advanced models that push the frontier of intelligence and capabilities, including new open source models,” Zuckerberg wrote in his announcement.
Internal benchmark data published by Meta indicates Muse Spark outperforms the latest leading models from OpenAI, Anthropic, Google, and xAI across a range of core tasks. “Muse Spark is the first step on our scaling ladder,” Meta shared in an official blog post, referencing the company’s long-term goal of building AI systems that far outstrip general human cognitive ability.
Artificial Analysis, an independent AI benchmarking firm that received early access to test Muse Spark, shared a positive assessment of the new model on social media, calling it one of the most capable systems the company has ever evaluated. “Muse Spark scores 52 on the Artificial Analysis Intelligence Index, placing it within the top 5 models we have benchmarked,” the firm said, noting its scoring rubric combines results from dozens of independent third-party performance tests.
Per Meta’s product overview, Muse Spark is natively multimodal, meaning it was trained from inception to process not just text, but also image, audio, and video inputs. The model also includes cutting-edge advanced reasoning capabilities—a defining feature of today’s highest-performing AI systems—and was built entirely from the ground up to deliver strong coding performance. Meta frames these core features as the foundation for building increasingly capable models using modern machine learning methods.
The company also designed Muse Spark to specialize in delivering reliable medical guidance. “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,” Meta explained in its blog post.
Since the underwhelming launch of Llama 4, Zuckerberg has poured massive resources into overhauling Meta’s entire AI division. The tech giant has poached top AI research and engineering talent from competing firms with compensation packages valued at hundreds of millions of dollars, and committed billions of dollars to acquiring or taking major stakes in a long list of emerging AI startups. Most notably, Meta recruited Alexandr Wang, CEO of leading AI training data firm Scale AI, to lead its overall AI efforts after investing $14.3 billion in the company.
Alongside the Muse Spark launch, Meta published a public document outlining its long-term vision for safely scaling AI models to superhuman levels of performance. The company’s new Advanced AI Scaling Framework lays out a formal set of safety checks that Meta will implement as its models grow increasingly powerful.
