In a significant announcement on Wednesday, Mark Zuckerberg revealed that Meta Superintelligence Labs has achieved a pivotal milestone with the introduction of a new family of AI models known as Muse, starting with the debut model named Spark. Zuckerberg shared this development through a Facebook post, stating that Muse Spark is now integrated into an updated version of Meta AI, accessible online at meta.ai or through the Meta AI app.
The announcement highlighted Muse Spark as a foundational step in scaling Meta's AI capabilities, described as a comprehensive overhaul of their AI initiatives. According to Meta, Spark is specifically designed to excel in everyday tasks that involve visual understanding, health management, shopping, and social interactions. Looking forward, Zuckerberg emphasized that Meta plans to create products that transcend mere question-and-answer functionalities, evolving towards AI that actively performs tasks on behalf of users.
Muse Spark: The First Major Offering from Meta Superintelligence Labs
This announcement marks the public unveiling of efforts that have been in progress since the summer of 2025, characterized by both ambition and challenges. In his manifesto published in July 2025, Zuckerberg articulated his vision for "personal superintelligence," aiming for an AI that assists individuals in achieving their personal goals rather than functioning under centralized control.
To realize this vision, Meta embarked on an extensive hiring initiative, recruiting over 50 leading researchers from prominent competitors such as OpenAI, Anthropic, and Google. The recruitment drive included bringing Alexandr Wang, the former head of Scale AI, to lead Meta's new superintelligence research division. However, shortly thereafter, Meta implemented a hiring freeze, citing standard budgetary processes, and restructured its team into four focused units: research, superintelligence development, products, and infrastructure. Zuckerberg justified this restructuring by asserting that significant advancements in AI are best achieved by compact teams that can maintain a comprehensive understanding of their projects, rather than larger, unwieldy organizations.
This rapid shift raised concerns amid wider market skepticism regarding the sustainability of the AI boom, especially following an MIT study indicating that most companies utilizing AI were not seeing a financial return on investment. In his manifesto, Zuckerberg made a clear distinction between Meta's approach and that of its competitors, asserting that while some AI labs aim to centralize superintelligence and deliver its output as a utility, Meta seeks to empower individuals.
In his recent announcement, Zuckerberg reiterated the lab's mission of "putting personal superintelligence in everyone's hands," advocating for the empowerment of individuals rather than the centralization of intelligence as the key to advancing humanity.
This new Muse Spark is the tangible result of Meta's multi-billion-dollar investments in AI development, with the company allocating $72 billion in 2025 and projected spending of up to $135 billion in 2026.
Muse Spark: Performance Benchmarks
Historically, Meta's Llama series of AI models has struggled to compete with its rivals in terms of performance metrics. While it remains to be seen whether Muse Spark can live up to the superintelligence label, this release marks a significant step forward for Meta following a period of hiring turmoil, team restructuring, and high-level conceptual discussions.
Zuckerberg expressed enthusiasm for future developments, stating, "I'm looking forward to sharing more soon." Alongside the Muse Spark announcement, Meta Superintelligence Labs disclosed its performance scores on various AI benchmark tests, including Humanity's Last Exam (HLE), ARC AGI 2, and GPQA Diamond. Although these scores could not yet be independently verified, Meta provided details regarding its testing methodology for Muse Spark.
Overall, Meta reported mixed outcomes when comparing Muse Spark to leading models such as Claude Opus 4.6 Max, Gemini 3.1 Pro High, GPT 5.4 Xhigh, and Grok 4.2, with Muse Spark outperforming some benchmarks while falling short in others. A comparative table of these benchmark performances was also released.
Accessing Muse Spark
Muse Spark is currently available to users. Those on desktop can access the AI model at meta.ai, while mobile users can try it within the Meta AI app. Additionally, Meta announced that select users would have the opportunity to access a private API preview.
To enhance its competitiveness against reasoning models developed by OpenAI, Anthropic, and Google, Meta plans to introduce a new "Contemplating" mode for Muse Spark, which will coordinate multiple agents to reason in parallel. This mode aims to provide significant improvements in complex task performance, achieving 58% in Humanity’s Last Exam and 38% in FrontierScience Research. However, this Contemplating mode is not currently available and will be rolled out gradually at meta.ai, with no specific timeline provided.
Source: Mashable News