Meta has introduced the Muse Wonder series, led by two models called Muse Wonder and Muse Wonder Full, which the company describes as its most capable models yet for personal and professional agentic tasks. According to Meta AI Research, the models deliver state-of-the-art performance on those tasks and combine two new capabilities, Active Memory and Active Focus, to keep track of what matters to a user and keep working toward their goals. Muse Wonder will first power experiences in Muse and roll out first to subscribers.
A new step on Meta’s scaling path
Meta AI Research says the release marks the next step on its scaling ladder and a major move toward what the company calls personal superintelligence. The company framed the work as a broader ambition rather than a single product update: it wants AI systems that can help individual people with a wide range of everyday and work-related goals.
The timeline matters to the story. Meta says it began a ground-up overhaul of its artificial intelligence efforts fourteen months ago, moving from rebuilding its technology stack from first principles to pushing the frontier of what its models can do. The Muse Wonder models build on the company’s earlier Muse Spark work, which outlined an approach to predictable and rigorous scaling. According to the post, the new models bring together work across pretraining and reinforcement learning, and were trained at Meta’s gigawatt-scale Prometheus data center.
Meta also says the models will support more than the chat experience. Muse Wonder will power artifact generation and other processes that require more intelligence. In the longer term, the company plans to use Muse Wonder to develop specialised models that it can bring to everyone using Muse.
What the models are designed to do
Domain expertise
Meta says Muse Wonder and Muse Wonder Full are meant to apply specialised knowledge to problems people face at work and at home. Examples include understanding a lease or making sense of financial or health information. The company’s evaluations include tasks in law, finance and health. One benchmark cited is Harvey’s Legal Agent Benchmark, known as LAB, which requires a model to work with source documents, follow detailed instructions and produce legal work products.
Active Memory and Active Focus
Meta identifies reliable memory as one of the most important steps toward useful personal and professional agents. Models need to remember what matters and recognise when earlier information becomes relevant again. Meta calls this capability Active Memory: carrying important context about a user across conversations and retrieving the details relevant to the task at hand.
Memory alone is not enough, according to the company. Completing a task also takes sustained effort. Active Focus describes how Muse Wonder keeps working toward a difficult goal by trying alternative approaches, working through blockers and adapting to new information. When needed, the model asks for clarification and checks in before taking important actions. Meta says it measures focus with a benchmark called FocusBench and measures sustained execution with SWE-Marathon, which tests work on long-running software projects.
Coding performance
Coding brings these abilities together, Meta argues. Building software involves tracking requirements, changing code across a codebase and repeating cycles of testing and revision. Meta says Muse Wonder and Muse Wonder Full deliver frontier coding performance for projects ranging from small applications to complex software, and that both were trained on diverse, long-running tasks to help them sustain progress.
Why breadth of ability matters
Meta’s post also addresses a debate familiar to anyone who follows AI closely. Two models can post similar benchmark scores, yet one may seem to handle a far wider range of unfamiliar tasks. Online critics have described the problem with the term “benchmaxxing,” meaning optimising a model for popular tests, while users praise what they call “big model smell,” a sense of broader general capability. Meta says both complaints point to a gap that a small set of well-known benchmarks cannot fully measure.
To study that gap, Meta compared performance on tasks it specifically targeted during post-training with performance on tasks it did not target. The company reports that both Muse Wonder Full and Muse Spark 1.3 improve in similar ways on the unfamiliar tasks as test-time compute increases. However, Meta says Muse Wonder Full performs better on unfamiliar tasks than the Muse Spark 1.3 trend would predict, given its results on the targeted tasks. The company attributes this to its larger, stronger base models, arguing that a stronger base lets post-training gains carry over more effectively to new situations.
This matters, Meta says, because personal agents must handle many combinations of goals, tools and constraints, and not all of them can be anticipated during training.
Safety and access
A personal agent can act on a user’s behalf and may handle sensitive material such as messages, calendars and private documents. Meta says this level of access requires a model that uses discretion with private information and resists misuse. The company says it trained the Muse Wonder family to lead on both points, including robustness to adversarial instructions and careful handling of private information.
Meta says Muse Wonder includes its strongest safeguards to date against chemical, biological, radiological and nuclear misuse, as well as cyber misuse. At the same time, the company says it is preserving access to frontier cyber capabilities for defensive use, in line with its Meta Superintelligence Scaling Framework. Meta describes Muse Wonder as its most aligned model to date. A full report with additional safety evaluations is due soon, and readers should look to that document for detail beyond the summary in the launch post.
What comes next
For now, the launch is significant mainly for its ambition and its emphasis on agents that remember, persist and check in with users. Meta has said the Muse Wonder models will reach subscribers first, and the safety report and any wider availability will be the next things to watch. The company’s own framing suggests that the real test will be how these agents perform on everyday legal, financial, health and software tasks once independent users begin to work with them.
Related: Helm Brings Shopify Store Monitoring Into One Dashboard · Pulse Local Keeps Your Budget Chat on Your Own Machine
Comments (0)