Meta has unveiled Muse Wonder and Muse Wonder Full, two new artificial intelligence models designed to handle complex personal and professional tasks, with a focus on long-term memory, sustained reasoning and advanced coding capabilities.
Announced on October 12, 2026, the models represent the latest stage in Meta’s effort to develop more capable AI systems that can work toward users’ goals across conversations, applications and extended workflows.
Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is currently accessible to Muse Code subscribers. Meta said Muse Wonder will soon power features within its Muse platform, beginning with subscribers, including artifact generation and tasks that require greater reasoning capabilities.
Building AI Agents That Remember and Follow Through
A central feature of the Muse Wonder series is what Meta calls Active Memory and Active Focus.
Active Memory enables the models to retain important information across conversations and retrieve relevant details when needed. The goal is to help AI agents understand a user’s preferences and circumstances rather than treating every interaction as an isolated request.
Active Focus, meanwhile, is designed to help the models work through challenging tasks over extended periods. This includes exploring alternative approaches, overcoming obstacles, adapting to new information and requesting clarification when necessary. Meta said the models are also designed to check with users before taking important actions.
These capabilities are intended to support a broad range of activities, from interpreting lease agreements and understanding financial or health information to planning personal schedules and managing complex professional workflows.
Meta evaluated the models across several benchmarks covering legal and financial tasks, long-term memory, business-process automation and software development. The company reported state-of-the-art results in key evaluations for personal and professional AI agents, although the published comparisons represent Meta’s own assessment of the models’ performance.
Stronger Coding and Performance on Unfamiliar Tasks
Mark Zuckerberg, chief executive officer of Meta Platforms Inc., reveals the Muse Charm device during the Meta Connect event in Menlo Park, California, US, on Wednesday, Sept. 23, 2026. Minh Connors | Bloomberg | Getty Images
Software development is another major focus of the release. Muse Wonder and Muse Wonder Full are designed to manage requirements across codebases, conduct repeated testing and revisions, and maintain progress on projects that require sustained effort.
Meta also highlighted the models’ ability to generalize beyond tasks specifically targeted during training. According to the company, Muse Wonder Full performs particularly well on unfamiliar tasks relative to the performance expected from its results on familiar evaluation tasks.
Meta attributes this advantage to stronger underlying models, which it says allow improvements gained during post-training to transfer more effectively to new situations. Such versatility could be important for AI agents that must respond to combinations of goals, tools and constraints that developers cannot anticipate in advance.
The company said both models were trained using its expanded AI infrastructure and advances in pretraining and reinforcement learning, building on the approach outlined with its earlier Muse Spark model.
Safety and Privacy Remain Central Concerns
Because personal AI agents may access messages, calendars and private documents, Meta said safety and responsible handling of information were central to the models’ development.
The company highlighted safeguards intended to resist adversarial instructions, prevent unauthorized code execution and reduce improper disclosure of personal information. It also said the models offer its strongest protections to date against chemical, biological, radiological and nuclear risks, as well as cyber misuse, while retaining capabilities that can support defensive cybersecurity work.
Meta described Muse Wonder as its most aligned model so far and said a more comprehensive report containing additional safety evaluations would be released later.
The launch marks another step in Meta’s ambition to bring more personalized AI capabilities to its products and users. Whether the models deliver consistently reliable results in everyday use will depend not only on benchmark performance but also on how effectively they handle real-world tasks, private information and decisions that require human oversight.
