Michael JohnsonA multi-step AI agent needs to maintain its goal, retrieve relevant information from prior context, and continue working after an error. On October 12, 2026, Meta introduced the Muse Wonder and Muse Wonder Full models, which the company designed for these scenarios. Muse Wonder is available to developers through the Meta Model API and Muse Code; Muse Wonder Full is available to Muse Code subscribers. Meta also says that Muse Wonder will later support Muse features, initially for subscribers.
The release is significant above all for its shift from demonstrations of personal assistants to models for developers. Meta describes tasks ranging from understanding documents and financial information to multi-step programming. Teams evaluating agent models can consult descriptions of the evaluations and a separate model card detailing comparison conditions and risk assessments.
What Meta means by Active Memory and Active Focus
In Meta’s description, Active Memory is the ability to carry relevant information across conversations and retrieve it when it helps with the current task. Active Focus refers to sustained work: the model tries alternative approaches, works through obstacles, and asks for clarification when needed. The company also says the model checks with the user before taking important actions.
These capabilities target tasks that require several steps to produce an answer, such as finding information in documents, planning, or making changes to a software project. Meta says it evaluates sustained performance using FocusBench and SWE-Marathon. The model card explains that SWE-Marathon evaluates large projects by whether tests pass, while FocusBench assesses how an agent overcomes obstacles and adapts to a changing environment across a series of interactions.
What is known about the evaluations
Meta published the Muse Wonder model card, which lists evaluations, methodology, and deployment conditions. According to the card, the models accept text, images, video, and audio, and have a stated inference context window of one million tokens. It describes evaluations of agent scenarios, programming, document handling, and safety.
Comparison conditions vary across evaluations. In controlled internal programming comparisons, Meta used a common testing setup; some other results came from external evaluators or model-provider reports. For example, SWE-Marathon results are based on data from Abundant AI: GPT‑6 Astra was evaluated through Codex, while Claude Fable 5.1 was evaluated through Claude Code. That row therefore does not represent a comparison of all models in a single testing setup.
Meta reports that Muse Wonder Full scored 92 flags in ExploitGym. The model card clarifies that 21 flags corresponded to suspected vulnerabilities, while the remaining 71 were obtained through unintended vulnerabilities. Only the first category counts toward the published ranking. This illustrates how evaluation methodology can change the meaning of a headline figure.
What to check before deployment
For a practical pilot, developers should start with tasks typical of their product: give the agent a limited set of tools, define a verifiable outcome, and measure the share of completed scenarios, the number of errors, and how often human intervention is needed. They should also test what happens with conflicting data, tool failures, and attempts to perform irreversible actions.
Meta’s model card recommends evaluating the model for its specific application and putting additional safeguards in place, including human confirmation for irreversible actions. It also notes that performance depends on the serving configuration, prompts, tools, and agent setup. The stated context window should therefore be considered alongside results on target tasks.
As of October 13, 2026, the release of both models, the stated access channels, and Meta’s published evaluation materials are confirmed. Availability through Muse for subscribers is a future step, not a launch already completed for all users. The release gives developers a basis for their own pilot; deployment decisions should be based on task results and safety checks in the target system.