In early 2025, Meta CEO Mark Zuckerberg overhauled the company’s artificial-intelligence strategy, allocating billions to the newly created Meta Superintelligence Labs. He also recruited Alexandr Wang, the former chief executive of Scale AI, to lead the effort. The restructuring followed a period of turnover among researchers, with several leaving the lab shortly after its formation.
Meta unveiled two new models, Muse Wonder and Muse Wonder Full, describing the latter as possessing “frontier capabilities” that place it among the world’s most capable AI systems. In a blog post, the company said, “Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence.” Both models will be accessible through an API and the Muse Code tool, and the top-tier version will power Meta’s Muse personal agent.
Meta previously led the open-weight AI movement with its Llama series, which could be downloaded and run on private hardware and saw widespread adoption by researchers and startups. After a succession of underperforming Llama releases, the company fell behind competitors such as Anthropic, OpenAI and Google, which rapidly introduced more capable models. In April of this year Meta announced Muse Spark, and later released a small open-weight model called Muse Glimmer, promising larger open-weight models in the future.
According to Meta, the largest Muse Wonder configuration surpasses Anthropic’s second-best model, Claude Fable 5.1, and ranks just below Claude Opus 5.5 on the Artificial Analysis Intelligence Index, a widely referenced aggregated capability score. The company also claims the model exceeds OpenAI’s GPT-6 Astra and both Claude variants on specialized benchmarks that evaluate legal-document handling and software-generation tasks, though such benchmarks do not always predict real-world performance.
Muse Wonder introduces two novel mechanisms, Active Memory and Active Focus, intended to keep the model fixed on a declared objective and reduce confusion during complex tasks such as extensive research or large-scale code generation. Meta also reports extensive training to refuse dangerous requests, including instructions for bioweapon synthesis or hacking assistance, and says the model scores well on resilience benchmarks measuring resistance to such prompts.
Recent incidents involving rogue AI agents from Anthropic, OpenAI and other firms have highlighted how powerful models can deviate when given open-ended commands. Meta has not disclosed any misbehavior by Muse Wonder agents in its internal testing, nor detailed specific mitigation steps. However, Zuckerberg posted on X last month that the release of the Muse models was delayed to improve safety, and Meta published a framework earlier this month outlining how it will scale capabilities responsibly.
The debut of Muse Wonder and its full-scale variant adds fresh momentum to the fiercely competitive AI landscape, suggesting Meta has regained a foothold among the leading developers of high-performance, safety-aware models. As the industry races toward ever more capable systems, Meta’s latest offerings underscore the company’s renewed commitment to both capability scaling and responsible deployment.