Muse Wonder and Muse Wonder Full are now generally available as frontier models tuned for personal and professional agents, combining Active Memory and Active Focus to track context, sustain effort, and complete complex tasks at work and in daily life.
What the Muse Wonder series delivers
The release introduces two models: Muse Wonder, accessible today through the Meta Model API and Muse Code, and Muse Wonder Full, available to Muse Code subscribers. Both were trained at the gigawatt-scale Prometheus data center and bring together pretraining and post-training advances to push capabilities further on agentic tasks.
Muse Wonder will soon power experiences in Muse, with subscribers receiving access first. It will also support artifact generation and other processes that require more intelligence. Specialized variants built on Muse Wonder are planned for broader release inside Muse over time.
Active Memory and Active Focus
Active Memory lets the model carry important context about a user across conversations and actively retrieve details relevant to the current task. Active Focus describes how Muse Wonder continues working toward a challenging goal: trying alternative approaches, working through blockers, adapting to new information, asking for clarification, and checking in before taking important actions.
FocusBench evaluates focus directly, while SWE-Marathon measures sustained execution on long-running software projects. Coding ties these capabilities together: building software requires tracking requirements, editing across a codebase, and cycling through repeated testing and revision. The Muse Wonder series supports that loop from small applications up to complex software.
Domain expertise across law, finance, and health
Muse Wonder and Muse Wonder Full apply specialized knowledge to problems people meet at work and in daily life, from reading a lease to making sense of financial or health information. Evaluations cover tasks in law, finance, and health. Harvey’s Legal Agent Benchmark (LAB), for example, requires models to work with source documents and follow detailed instructions to produce legal work products.
Capabilities that generalize beyond training targets
Models with similar benchmark scores can still differ sharply on unfamiliar tasks, a gap sometimes framed as “big model smell” versus “benchmaxxing.” Scaling test-time compute reveals the difference: comparing Muse Spark 1.3 with Muse Wonder Full, both improve on out-of-distribution tasks as compute rises, but Muse Wonder Full sits above the trend line. The reason is a stronger base model, since gains from post-training carry over more effectively to unfamiliar tasks when the foundation is larger and more capable.
That breadth matters for personal agents, which must handle combinations of goals, tools, and constraints that are not all anticipated during training. Demo examples include building a running plan from personal preferences, designing custom swag from a real photo collection, and preparing a custom wedding seating chart from a complex list of matching requirements.
Safety at the frontier
A personal agent that handles messages, calendars, and private documents needs a model that exercises discretion with private information and resists misuse. The Muse Wonder family is trained to lead on robustness to adversarial instructions and careful handling of private information. Safeguards are tuned against CBRN (chemical, biological, radiological, and nuclear) and cyber misuse, while keeping frontier cyber capabilities available for defensive use, within the Meta Superintelligence Scaling Framework. A full report with additional safety evaluations is planned.
Where Muse Wonder is available
Muse Wonder is available today through the Meta Model API and Muse Code, with Muse Wonder Full available to Muse Code subscribers. The models will soon begin powering experiences in Muse, starting with subscribers, and will support artifact generation and other intelligence-heavy processes. Future specialized models built on Muse Wonder are planned for broader use inside Muse.
FAQ
What is Muse Wonder?
Muse Wonder is a frontier model in the Muse Wonder series, alongside Muse Wonder Full. It is built for personal and professional agents and combines Active Memory with Active Focus to handle real-world tasks at work and in daily life.
Where can I access Muse Wonder?
Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers. Both models will soon begin powering experiences in Muse, with subscribers gaining access first.
What benchmarks and capabilities does Muse Wonder lead on?
The series targets personal and professional agent tasks, memory, and coding, with evaluations including Harvey’s Legal Agent Benchmark (LAB) for law, plus tasks in finance and health, and FocusBench and SWE-Marathon for sustained execution.
Related coverage
This article summarizes reporting from research.meta.ai. See our editorial disclaimer for how our articles are produced.
Run a free scan to see your AI Visibility Score, SEO rating, and local citation accuracy.