- Meta launches Muse Wonder and Muse Wonder Full models, boosting complex task handling capabilities.
- Muse Wonder integrated into consumer assistant, initially limited to subscribers.
- Active Memory feature retains relevant information in conversations, improving user experience.
- Meta emphasizes safety measures for handling private information, resisting malicious instructions.
Meta has launched two new frontier models designed to leverage long-term memory, sustained reasoning, and software tools to execute complex personal and professional tasks. The company says Muse Wonder and Muse Wonder Full demonstrate state-of-the-art performance across evaluations covering business automation, programming, contextual memory, and professional work. The announcement did not provide specific benchmark scores, making it difficult for outsiders to independently assess the magnitude of the performance gains. Muse Wonder is available via the Meta Model API and Muse Code, while Muse Wonder Full is offered to Muse Code subscribers.
Plans to Integrate into Consumer Assistant
Meta plans to integrate Muse Wonder into its consumer assistant, initially limiting access to subscribers for generating digital works and handling tasks requiring more advanced reasoning. This release follows Meta's launch of Muse Spark and comes 14 months after the company rebuilt its AI infrastructure. The new models were trained in its kilowatt-scale Prometheus data centers, reflecting Meta's investment in larger compute systems and more powerful models.
A core feature of Muse Wonder is Active Memory, which enables the model to retain relevant information during conversations and retrieve it when needed. The system is designed so users do not need to repeat prior instructions; instead, it recognizes the importance of previously shared details for new tasks. This could help personal assistants handle work involving preferences, documents, and ongoing projects. We are excited to introduce the Muse Wonder family, including Muse Wonder and Muse Wonder Full, our most powerful models to date.
Both are frontier models designed to deliver state-of-the-art performance for personal and professional agentic tasks.
Active Focus Enhances Model Operational Capability
Meta also introduced Active Focus, a capability designed to let models sustain operation when facing difficult tasks. The system can try different approaches, cope with unexpected issues, and seek clarification when necessary. It is also designed to request confirmation before executing important actions. The company evaluated contextual memory using MemoryBench and sustained task execution using FocusBench. Additionally, it used AutomationBench-AA to assess business workflows and professional evaluations covering legal and financial tasks.
Focus on Long-Term Software Development
Software development is another key focus of the Muse Wonder family. Meta says both models can handle long-horizon programming tasks involving multiple files, changing requirements, and repeated testing. The company used SWE-Marathon and DeepSWE to evaluate performance on long-horizon software engineering tasks. These evaluations check whether models can maintain progress in complex projects rather than merely generating isolated code snippets.
Meta also tested its models on unfamiliar tasks. According to the company, Muse Wonder Full demonstrated generalization ability exceeding the performance trend established by Muse Spark 1.3. This suggests improvements may extend beyond the specific tasks emphasized during training, though the announcement did not include sufficient numerical results to determine the size of that advantage. Meta demonstrated Muse Wonder creating a playable 3D kitchen game from image input and building a personalized fitness plan through browser and app interactions.
The company also highlighted safety measures for handling private information, resisting malicious instructions, and limiting dangerous uses involving biological and cybersecurity threats.
Meta plans to use Muse Wonder to develop specialized models for its broader product ecosystem, though no specific date for wider consumer availability has been announced.
Impact of the Muse Wonder Family on AI
Meta's launch of the Muse Wonder family of models demonstrates its continued innovation and investment in artificial intelligence. These models not only enhance complex task handling capabilities but also introduce new features such as Active Memory and Active Focus, improving the fluidity of user interactions. With the release of these models, Meta further integrates AI technology into its product ecosystem, aiming to elevate the consumer experience. The emphasis on safety measures also underscores the company's commitment to data protection, which is especially critical in today's digital environment.