Meta unveils Muse Wonder series for personal and professional AI agents

Meta released its Muse Wonder model series on October 12, 2026, introducing Muse Wonder and Muse Wonder Full as frontier models built for personal and professional agentic tasks. The models combine what the company calls Active Memory and Active Focus to track user context and sustain effort on complex goals, marking a push toward what Meta describes as personal superintelligence.

Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is available for Muse Code subscribers. The models will soon power experiences inside Muse, rolling out first to subscribers for artifact generation and processes requiring more intelligence.

Scaling from first principles

Fourteen months ago, Meta began rebuilding its AI stack from scratch. The Muse Wonder release builds on that work across pretraining and reinforcement learning. Both models were trained at Meta's gigawatt-scale Prometheus data center.

"When we introduced Muse Spark, we outlined our approach to predictable, rigorous scaling," the company said. "Today's release builds on that work across pretraining and reinforcement learning, delivering our most capable models yet."

Active Memory and Active Focus

The models introduce two capabilities central to agentic work. Active Memory allows the model to carry context about a user across conversations and retrieve relevant details when earlier information becomes useful. Active Focus describes sustained effort toward a goal - trying alternative approaches, working through blockers, and asking for clarification before taking important actions.

Meta evaluates focus through FocusBench and sustained execution through SWE-Marathon, which tests work on long-running software projects. The company said both models were trained on diverse, long-running tasks to help them sustain progress across repeated cycles of testing and revision.

Performance that generalizes

Meta also addressed the gap between benchmark scores and real-world versatility. The company compared Muse Spark 1.3 and Muse Wonder Full on unfamiliar tasks as test-time compute scaled. While both models improved similarly, Muse Wonder Full outperformed what the Spark trend would predict - a result Meta attributes to larger, stronger base models whose post-training gains carry over more effectively to tasks not targeted during training.

"That versatility especially matters for personal agents, which need to handle many combinations of goals, tools, and constraints - not all of which we can anticipate during training," the company said.

Safety and alignment

Because personal agents may handle messages, calendars, and private documents, Meta trained the Muse Wonder family for discretion with private information and resistance to adversarial instructions. The models offer what Meta calls its strongest safeguards to date against CBRN and cyber misuse, while preserving access to frontier cyber capabilities for defensive use under the company's Superintelligence Scaling Framework.

"Muse Wonder is our most aligned model to date, and alignment is central to how we build personal superintelligence," Meta said. A full safety report will follow.

Why this matters for executives and product teams

Muse Wonder signals a shift from general-purpose models toward AI that maintains context about specific users and their goals over time. For product development and strategy leaders, the Active Memory architecture points to a future where AI tools don't just answer prompts - they remember project history, track requirements across conversations, and work through multi-step tasks without losing the thread. The generalization findings also matter: a model that handles unfamiliar combinations of goals and constraints reduces the risk of brittle automation in customer-facing or internal workflows. Teams evaluating AI Agent Courses or Generative AI Courses should watch how these memory and focus capabilities reshape what agents can reliably do in production.