Meta announced Muse Wonder and Muse Wonder Full on October 12, 2026, positioning the models for AI agents at work and in everyday life. Muse Wonder is available through the Meta Model API and Muse Code, while Wonder Full is offered to Muse Code subscribers. Wider integration into Muse remains planned for later, starting with subscribers.
A task does not always end with a single answer. An assistant needs to remember earlier decisions, continue after a failure and know when to stop and ask. Meta presents Active Memory and Active Focus as foundations for that shift. These are the company's descriptions of capabilities, not promises that every workflow will finish without errors or oversight.
Who can access the models now
Muse Wonder's immediate entry point is primarily the development environment: an API for integrating the model into a product, and Muse Code for those using the company's coding tool. Muse Wonder Full is available to Muse Code subscribers. These access routes should be kept distinct.
Meta says Muse Wonder will soon power more experiences within Muse, initially rolling out to subscribers. It also mentions artifact generation and processes requiring greater capability, followed by specialized models for a wider audience. These are future plans, rather than availability to every Meta AI user today.
A developer can examine the announced development route and its terms, while someone seeking a ready-made personal assistant needs to check what is actually offered in their account. There is no reason to change a work plan merely because a capability appeared in a demonstration if it cannot yet be accessed in the product being used.
The announcement does not separately confirm every feature in Israel or provide a dedicated assessment of Hebrew performance. A general description of a personal assistant does not establish suitability for every local document or Hebrew conversation. An Israeli team can test its own language tasks when the relevant route becomes available, without assuming the launch has already settled that question.
Active Memory is meant to retain what matters
Meta defines Active Memory as keeping important context across conversations and retrieving earlier details when useful to the current task. The point is more than storing additional history. A useful assistant should distinguish what remains relevant, what has changed and what does not belong in the request.
A planning example makes the value clearer: in an earlier conversation, a team set its audience, budget and elements that must not change. At the next meeting, it asks for the deliverable to be revised. Useful memory avoids entering those decisions again, but its value also depends on recognizing a new correction.
Relevant information from a previous conversation is not automatically authorized for sharing with another person or an external service. An organization should define in advance which details may move between tasks, who may use them and which actions require further approval. Memory is no substitute for access policies and clear ownership of information.
Before broad adoption, a simple test is possible: give the system several decisions, correct one and request a new task that depends on the correction. This tests not only recall, but also whether it prioritizes current information and can explain the basis for its answer.
Active Focus aims to continue when something goes wrong
In Meta's description, Active Focus means persisting toward a difficult goal by trying alternatives, recovering from errors and adapting to new information. The company says the model should ask for clarification when needed and check in before important actions. The intended outcome is not blind persistence, but continued work that respects the user's intent and decision points.
A seemingly successful code fix can fail a test, and changing one file may require changes elsewhere. Meta says both models were trained on varied long-running tasks to support that process. Code quality should still be determined by the specific project's tests, requirements and review, rather than how long an agent kept working.
Meta's research blog names FocusBench for evaluating focus and SWE-Marathon for work on sustained software projects. Even when a company reports leading results, its testing conditions differ from another team's repository, permissions and tools.
A practical trial can start with a task that is stoppable and reversible, with clear success criteria. Examine how the agent responds to missing information, a failed test or changed requirements, and how much intervention brings it back on track. Action count alone is insufficient: did persistence move it toward a correct result within the agreed boundaries?
Greater capability also needs clear boundaries
Meta also positions the models in fields such as law, finance and health, as tools for understanding information and completing tasks. That does not authorize treating model output as an unchecked professional decision. An organization might assess whether an agent found specified clauses in a document; a consequential judgment still needs the qualified person and appropriate context.
The company says it trained the Muse Wonder family to avoid disclosing private information without permission, refuse assistance with harmful acts and preserve defensive cyber uses. It links the launch to the Meta Superintelligence Scaling Framework, its framework for evaluation and safety requirements. These describe development and protection practices, rather than proving every real-world risk has been resolved.
Meta says a full report with additional safety evaluations will follow. The safeguards already described should therefore be distinguished from details not yet published in the launch announcement. Organizations can start with nonsensitive material and defined tasks, assessing outcomes before broadening permissions.
Meta's move sets a more useful standard: does the assistant remember the right decision, change course after failure and ask for approval at the right point? Muse Wonder and Wonder Full are the company's answer to that challenge. Their value will be established when those capabilities meet users' actual work, access conditions and the checks needed to trust the result.