Meta has introduced Muse Wonder and Muse Wonder Full, two AI models for personal and professional agentic tasks. The models build on Muse Spark and Meta’s work on pretraining and reinforcement learning as part of the company’s broader overhaul of its AI efforts.

Meta Muse Wonder and Muse Wonder Full

Meta describes Muse Wonder and Muse Wonder Full as its most capable models to date. The company said it began overhauling its AI efforts 14 months ago, moving from rebuilding its foundations to developing models for more complex tasks.

When Meta introduced Muse Spark, it outlined an approach to predictable, rigorous scaling. The Muse Wonder series builds on that work across pretraining and reinforcement learning. Both models were trained at Meta’s gigawatt-scale Prometheus data centre.

Meta’s benchmark evaluations cover personal and professional agents, memory, coding, and general capabilities. The company’s comparison also includes other frontier models across personal and professional agent tasks, coding, and general evaluations.

Domain expertise and coding

Muse Wonder and Muse Wonder Full handle tasks involving specialised knowledge in law, finance, and health. Meta’s evaluations include Harvey’s Legal Agent Benchmark (LAB), which tests whether models can work with source documents and follow detailed instructions to produce legal work products.

For coding, Meta focused on tasks that require models to track requirements, make changes across a codebase, and test and revise software. Both models were trained on diverse, long-running tasks to support development ranging from small applications to complex software projects.

Active Memory and Active Focus

The models include two capabilities intended to help them retain context and make progress on tasks over time.

• Active Memory: Retains important context across conversations and retrieves relevant details when they are needed for a task.
• Active Focus: Helps the models continue working towards a goal by trying alternative approaches, working through obstacles, recovering from errors, and adapting to new information. They can also ask for clarification when necessary and check with users before taking important actions.

Meta evaluates focus using FocusBench and sustained execution using SWE-Marathon, which tests models on long-running software projects.

Generalisation to unfamiliar tasks

Meta examined how well the models perform on tasks that were not specifically targeted during post-training. The company discussed this in relation to “benchmaxxing,” which refers to optimising performance on selected benchmarks, and “big model smell,” which describes broader capabilities that may not be fully reflected in scores across a limited set of evaluations.

To assess generalisation, Meta compared Muse Spark 1.3 and Muse Wonder Full at different levels of test-time compute. Each point in the comparison represents a distinct reasoning effort. The evaluation uses two measures:

• In-distribution performance: The horizontal axis shows a composite score for tasks targeted during post-training.
• Out-of-distribution performance: The vertical axis shows the average score for tasks that were not targeted during post-training.

Meta said both models improve similarly on out-of-distribution tasks as test-time compute increases. However, Muse Wonder Full performs better on unfamiliar tasks than the trend for Muse Spark 1.3 would predict based on its in-distribution performance.

The company attributes this difference to the larger, stronger base models underlying the Muse Wonder series. According to Meta, a stronger base model allows improvements from post-training to transfer more effectively to unfamiliar tasks. This is relevant to personal agents because they must handle combinations of goals, tools, and constraints that cannot all be anticipated during training.

Safety and privacy

Personal agents may access messages, calendars, and private documents. Meta said it trained the Muse Wonder family to protect private information, resist misuse, and support defensive security work. The models are trained to:

• Avoid sharing private information without permission.
• Refuse requests to assist in harmful acts.
• Resist adversarial instructions.
• Help developers and security teams identify vulnerabilities in their own code and strengthen their systems.

Meta said the models have its strongest safeguards to date against chemical, biological, radiological, and nuclear (CBRN) misuse and cyber misuse, while retaining cybersecurity capabilities for defensive use.

The company trains, evaluates, and releases its most capable AI models under the standards set out in its Meta Superintelligence Scaling Framework. Meta describes Muse Wonder as its most aligned model to date. A full report containing additional safety evaluations will be available soon.

Availability

Muse Wonder is available to developers and builders through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers.

Meta plans to bring Muse Wonder to more experiences in its Muse app, starting with subscribers. The model will also support artifact generation and processes that require more intelligence. Meta plans to use it to develop specialised models for future Muse experiences.