Muse Wonder: Advancing the Frontier of Professional AI Agents

---

Date: October 12, 2026 | Reading Time: ~6 minutes

Introducing a New Era of Agentic AI

We are thrilled to unveil the Muse Wonder series, featuring Muse Wonder and Muse Wonder Full. These represent our most sophisticated models to date, specifically engineered to provide state-of-the-art performance for professional and personal agentic workflows.

By integrating Active Memory and Active Focus, these models don't just respond—they remember what is important to the user and persevere through complex objectives. This release represents a significant milestone on our scaling trajectory, moving us closer to the realization of personal superintelligence.

🚀 Availability & Access

• Meta Model API: Muse Wonder is available now.
• Muse Code: Both Muse Wonder and Muse Wonder Full (exclusive to subscribers) are live.
• Muse App: Integration is coming soon, rolling out first to subscribers.
• Advanced Features: These models will soon drive artifact creation and high-intelligence processes.
• Future Outlook: We will leverage Muse Wonder to synthesize specialized models for the broader Muse community.

---

The Path to Personal Superintelligence

Our current success is the result of a comprehensive, 14-month architectural overhaul. We transitioned from rebuilding our entire stack based on first principles to pushing the absolute boundaries of the frontier.

Building upon the foundation of Muse Spark and our commitment to predictable, rigorous scaling, we utilized the Prometheus data center—a gigawatt-scale facility—to train these models. This effort combined cross-functional expertise from across Meta to ensure these tools can handle the intricacies of both professional labor and daily life.

Core Architecture Logic

---

Frontier Capabilities & Model Highlights

The Muse Wonder series sets new benchmarks in memory, coding, and agentic behavior. While general benchmarks provide a high-level overview, the true power lies in four specific pillars:

1. Domain Expertise

The models possess deep specialized knowledge, allowing them to navigate complex real-world documentation.

• Legal: Capable of producing professional work products using source documents (tested via Harvey’s Legal Agent Benchmark).
• Finance & Health: Ability to synthesize and explain intricate financial statements or medical data.

2. Active Memory

Reliable memory is the bedrock of a useful agent. We define Active Memory as the ability to:

• Maintain critical user context across multiple disparate conversations.
• Dynamically retrieve specific details exactly when they become relevant to the current task.

3. Active Focus

Completing a difficult goal requires more than a single prompt; it requires sustained effort. Active Focus allows the model to:

• Iterate through alternative strategies when blocked.
• Adapt to new information in real-time.
• $\text{Proactively ask for clarification} \rightarrow \text{Verify before taking critical actions}$.

4. Frontier Coding

From simple scripts to massive software architectures, Muse Wonder excels in long-term development.

• Evaluation: Tested via FocusBench and SWE-Marathon.
• Workflow: The models can track requirements, modify large codebases, and engage in continuous cycles of $\text{testing} \rightarrow \text{revision}$.

---

Generalization: Beyond "Benchmaxxing"

In the AI industry, there is often a gap between benchmark scores and real-world utility. Some call the pursuit of high scores "benchmaxxing", while the intuitive feeling of a truly capable model is often described as "big model smell."

To measure this, we analyzed how performance on targeted post-training tasks ("in-distribution") correlates with performance on unfamiliar tasks ("out-of-distribution") as we increase test-time compute.

"A stronger base model ensures that the gains achieved during post-training are not just memorized patterns, but capabilities that generalize to the unknown."

The Scaling Relationship

We can express the relationship between reasoning effort ($E$) and performance ($P$) as: $P_{out} \approx f(P_{in}, E)$

Our data shows that while both Muse Spark 1.3 and Muse Wonder Full improve as compute scales, Muse Wonder Full exhibits a significant upward shift. It performs better on unfamiliar tasks than its in-distribution performance would suggest, proving that a more powerful base model yields more effective generalization.

Implementation Example

To utilize the agentic focus of Muse Wonder Full, a developer might use a configuration similar to this:

{
  "model": "muse-wonder-full",
  "agent_settings": {
    "active_memory": true,
    "active_focus": "high",
    "verification_threshold": 0.85,
    "max_iterations": 10
  }
}

Final Checklist of Goals

• Rebuild stack from first principles
• Scale via Prometheus Data Center
• Implement Active Memory & Focus
• Deploy specialized models to all Muse users