Meta’s Muse Wonder models target personal and professional agents

The release is about sustained work, not a new chatbot name

Meta announced Muse Wonder and Muse Wonder Full on October 12. The company presented them as its most capable models for personal and professional agent work, with examples ranging from sorting through an insurance claim to understanding a lease or building an app. Meta’s central pitch is that a model should keep working when a task takes more than one exchange: remember useful context, try another approach after a failure, and know when to ask a person a question.

Meta calls those ideas Active Memory and Active Focus. Those labels describe capabilities the company says it trained into the models; they do not establish that the system remembers every detail correctly or finishes every multi-step task. TechCrunch’s report describes the launch and notes that the models’ benchmark figures came from Meta’s own testing. The article is useful for the release context, but it is not an independent head-to-head test of everyday agent work.

A long task makes model quality only one part of the result

Meta’s current developer guide describes Wonder as handling multi-file coding, planning, delegating to sub-agents, calling tools in parallel, and recovering from errors. It accepts text, images, video, audio, and PDFs, then returns text. That is a wider interface than a text-only prompt, but the model does not act on a file or account by itself: the surrounding product still determines which tools it can reach and what permissions a user has granted.

This is where an agent can feel more capable than a conventional chat window. A coding assistant may inspect a repository, change several files, run checks, and revise its work. A personal agent may bring together prior context while carrying out a request. In both cases, a useful result depends on the environment, the available tools, and the user’s ability to see what happened. A benchmark score for a model alone cannot tell a buyer how safe or reliable a particular application will be.

Meta’s developer page lists a 1,048,576-token context window. That describes how much input can fit in a request; it does not mean an assistant permanently remembers all of it or that a user can later recover every source. Active Memory is the separate cross-conversation feature Meta describes. A team should test what survives between tasks and how to correct it, rather than assume a large context limit guarantees perfect recall.

The current access map is more specific than launch-day wording

At launch, Meta said Wonder was available through the Meta Model API and Muse Code, while Wonder Full was available to Muse Code subscribers. The company also said Wonder would soon power more experiences in the consumer Muse product, starting with subscribers. Meta’s current Model API documentation now says Wonder is available on the API and in Muse Code, and that Wonder Full is available only in Muse Code. The consumer-app sentence in the launch notice still gives no date or rollout schedule.

The documentation sets a sharper boundary than the shared family name suggests. The standard Wonder model is available through the API in Standard mode and through every Muse Code plan. Wonder Full is not available through the API; Muse Code subscribers need the Power Usage or Maximum Usage plan to use it. Choosing the highest reasoning effort for standard Wonder does not switch it to Wonder Full. Meta also says these models are not currently available in South Korea. Those are practical limits a developer should check before planning around a model identifier or subscription.

Meta’s product hub now features Wonder among its latest models, but that label by itself does not establish that every consumer Muse account has received the model. A user looking for the personal assistant should check the model shown in their own account rather than infer availability from a developer launch or a site-wide product tile.

Muse Code’s data setting deserves a separate check

The current developer documentation also distinguishes two content modes for Muse Code. In Standard mode, Meta says prompts, code, files, context, and responses are not used to train Meta models. Contributor mode gives subscribers more usage, but the documentation says it is enabled by default and permits Meta to use content to train future models; users can turn it off in settings. For the pay-as-you-go API, Contributor mode is unavailable.

That is a product-setting detail, not a general statement about every Meta AI service or account. A team handling private code or customer records should confirm the mode on the workspace it will actually use, along with its retention and access terms. It should not assume that a model’s availability page also describes the organization’s complete privacy arrangement.

Memory is not the same as a saved transcript

Persistent memory changes what a person needs to review. A user may intentionally state a preference, or an agent may retrieve an earlier detail while handling a new task. It helps to see whether a recalled fact came from a supplied document, an earlier conversation, or a connected service. Meta describes Wonder as carrying relevant context between conversations, but the launch announcement does not spell out every control a person has over each remembered detail.

Meta’s current Help Center says a user can ask Muse what it has stored and can view, edit, or delete the Memory file. The same guide cautions that Muse may still remember information it learned from material after that material is deleted. That offers a management path, but it is not a promise that every trace is erased. Before adding durable work or personal context, users should check what the assistant has retained and correct stale information.

Meta describes safety goals, while the additional report remains pending

Meta says the Wonder family was trained to avoid disclosing private information without permission, refuse requests to assist with harmful acts, and strengthen security against cyber attacks while preserving defensive security work. It says the models were evaluated under the Meta Superintelligence Scaling Framework. The accompanying research post gives more detail about the company’s tests and reported results, but it is still a company-authored account of its own evaluation.

The announcement says that a full report with additional safety evaluations will be available soon. In the official material reviewed on October 23, that wording remains the published status. That does not mean no testing occurred; it means readers still lack the additional public report Meta itself promised. A buyer evaluating an agent for sensitive files or customer messages should ask which controls belong to the model, which belong to the product, and what evidence supports each one.

Meta’s own description gives a sensible set of questions for a pilot: Does the agent surface the sources behind a remembered detail? Can a person stop or correct a long-running task? Does it ask before taking a consequential action? Can the administrator review the activity afterward? The release explains what the models are designed to do, while answers to those deployment questions depend on the service and its settings.