Meta has unveiled the Muse Wonder series, introducing Muse Wonder and Muse Wonder Full as its most capable AI models designed for personal and professional agentic tasks. According to Meta’s research blog, the series centers on two new mechanisms: Active Memory, which retains and retrieves relevant user context across conversations, and Active Focus, which sustains effort toward challenging goals by trying alternative approaches, adapting to new information, and seeking clarification when needed. These capabilities are intended to help the models keep track of what matters to users and work diligently toward their objectives.
This image illustrates the branding of the Muse Wonder series introduced in the article, supporting the description of the new models.
The Muse Wonder family demonstrates state‑of‑the‑art performance on a range of benchmarks that measure domain expertise, memory, and coding. Meta reports strong results on tests of legal document understanding (using Harvey’s Legal Agent Benchmark), financial and health reasoning, and frontier coding tasks that involve building software from small applications to complex codebases. Additionally, when evaluating generalization beyond the data seen during post‑training, Muse Wonder Full shows improved performance on unfamiliar tasks compared to earlier models like Muse Spark 1.3 when test‑time compute is scaled, indicating that its stronger base model helps transfer learned abilities to novel situations involving varied goals, tools, and constraints.
Safety and alignment are highlighted as key considerations for the Muse Wonder series, given the potential for personal agents to access sensitive information. Meta states that the models have been trained to lead on robustness against adversarial instructions and careful handling of private data, offering what it describes as the company’s strongest safeguards to date against CBRN (chemical, biological, radiological, and nuclear) and cyber misuse while preserving defensive cyber capabilities in line with its Meta Superintelligence Scaling Framework. The models are available today through the Meta Model API and Muse Code, with Muse Wonder Full accessible to Muse Code subscribers, and are slated to power experiences in the Muse ecosystem, artifact generation, and other processes requiring higher intelligence.