Meta Releases Muse Wonder Models for Agent Work
Muse Wonder is available through Meta's developer channels while consumer rollout and fuller safety evidence remain later checkpoints.
Edited by Tyronne Panaino
Meta released Muse Wonder and Muse Wonder Full on October 12 as models for personal and professional agent tasks. The Meta research post says Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers.
The release matters to developers building agents that must carry context across conversations and continue working through multi-step tasks. It does not yet represent a broad consumer launch: Meta's newsroom announcement says Muse Wonder will come to Muse later, beginning with subscribers.
Availability differs across the two models
Muse Wonder is the wider developer release. Meta lists both its model API and Muse Code as current access points. Muse Wonder Full has the narrower path, limited in the fetched evidence to Muse Code subscribers. Meta does not describe Full as generally available through the model API.
That distinction is useful for teams evaluating the launch. A model name in an announcement is not the same as identical access across every product surface. The fetched pages also do not provide a complete public schedule for the consumer rollout, so the present checkpoint is developer availability rather than universal access inside Muse.
Active Memory and Active Focus define the agent design
Meta describes Active Memory as a way to preserve important context across conversations and retrieve details when they become relevant. Active Focus is intended to keep the model working toward a difficult goal by trying alternatives, recovering from errors and asking for clarification. Meta also says the system checks in before important actions.
Together, those features move the product pitch beyond a single prompt and response. They target agents that need continuity and persistence, which also makes user control and private-context handling more important. Meta says the models were trained to handle private information carefully and resist adversarial instructions, but those safeguards remain company-reported in the evidence available for this release.
Meta links the models to work, coding and sustained execution
The research post says Meta evaluates the models on tasks in law, finance and health, as well as memory and coding work. It names FocusBench for focus and SWE-Marathon for sustained execution on long-running software projects. Meta also says the models were trained at its Prometheus data center and build on the earlier Muse Spark work.
These descriptions establish the intended evaluation scope, not independent performance. The public pages make broad comparative claims but do not provide third-party reproduction in the evidence used here. They also do not establish reliability for a particular legal, financial, health or software workflow. Buyers and developers still need tests that reflect their own tools, permissions, data boundaries and review requirements.
The safety report is still a missing checkpoint
Meta says Muse Wonder has stronger safeguards for private information, adversarial instructions, chemical and biological misuse, and cyber misuse. It also says defensive cyber work remains available under its scaling framework. However, the company says a fuller report with additional safety evaluations will be published later.
That leaves an important evidence gap at launch. The fetched material does not yet show the detailed evaluation results, failure rates, red-team coverage or independent review needed to judge how the safeguards behave under sustained agent use. It also does not publish a complete account of regional availability, model pricing, context limits or operational rate limits.
What developers should verify
A useful evaluation should separate model capability from product access. Teams can confirm whether their account exposes Muse Wonder through the Meta Model API or Muse Code, then test memory retrieval, task persistence, permission boundaries and intervention behavior with representative workflows. Muse Wonder Full should be treated as a distinct subscriber offering until Meta documents broader access.
The next verifiable checkpoints are the promised safety report, concrete consumer-rollout dates and operational documentation for the developer endpoints. Those updates will show whether Meta's agent design translates from launch claims into repeatable behavior across real deployments.
Status
Confirmed. Meta has released Muse Wonder through its model API and Muse Code and Muse Wonder Full through Muse Code subscriptions. Internal confidence is medium because capability and safeguard claims are first-party, the consumer rollout is still pending and the detailed safety report is not yet public.
Sources
Update note: Last reviewed 2026-10-13. We will revise this post when Meta publishes the detailed safety report or changes model availability.
Sources
- Meta Newsroom — Introducing Muse Wonder and Muse Wonder Full — official
- Meta AI Research — Muse Wonder — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.