Meta introduced Muse Wonder and Muse Wonder Full on October 12, extending its model family for personal and professional agents. The release emphasizes memory across conversations and sustained work on difficult tasks. For teams already using Muse Code, the immediate question is whether either model can finish more of a real project with fewer interventions. Meta’s announcement
What is available now
Meta lists Muse Wonder as available through the Meta Model API and Muse Code. Wonder Full is available to Muse Code subscribers. The announcement says Wonder is coming to the consumer Muse product soon, beginning with subscribers; that rollout is separate from today’s developer availability. Release and access details
That distinction matters when planning an evaluation. Access to Wonder in an API does not establish equivalent API access to Full. Before building around either model, check the account’s available model identifiers, limits and current pricing. This article does not assume downloadable weights or a self-hosting option.
The model card lists a one-million-token inference context and text, image, video and audio inputs, with text output. Model capability and product support are separate: verify the modalities exposed by the interface you plan to use.
What Meta means by memory and focus
Meta uses Active Memory for carrying important context across conversations and retrieving relevant details. Active Focus describes sustained work toward a goal, including trying alternatives and adapting when new information arrives. Meta says the models were trained on long-running tasks and reports evaluations covering agents, memory and coding. These are the developer’s results, not S5 Labs measurements. Capabilities and evaluation discussion
For a useful trial, give the agent a small repository task with explicit acceptance criteria. Introduce a revised requirement partway through, then check whether it updates the implementation and tests without retaining the superseded instruction. In a separate session, test whether the relevant project constraint is recalled accurately. Record the corrections you had to make, not just whether the final demonstration looks convincing.
Memory deserves its own acceptance criteria. A remembered preference may be useful; an outdated deployment target can cause a costly mistake. Check whether users can inspect and correct retained information, and whether the agent distinguishes a durable instruction from a temporary exception. Those are evaluation questions, not claims that every Muse surface exposes the same controls.
Benchmark settings matter
Meta’s model card says its Wonder results use maximum reasoning effort. Its comparisons combine controlled internal runs with externally reported results, and external coding scores retain their original harness or agent product. The card also describes competitor configurations as best-effort where Meta ran them internally.
Use those qualifications when reading a leaderboard gap. An evaluation at maximum reasoning effort does not tell you the latency or cost of the setting you will deploy. Compare the complete model-and-tool configuration at a budget you can sustain.
Assess the complete workflow
The launch follows Muse Spark 1.3 and the earlier Muse Code release. Compare Wonder against the workflow your team actually uses, including repository access, tool permissions and review steps. A model-only score cannot tell you how much unattended work your integration should permit.
Our assessment: start with a bounded task that produces a reviewable artifact, then compare completion quality, elapsed time and human correction effort. Keep publication, deployment and external communication behind explicit approval while evaluating. A more persistent agent is useful when it preserves the right constraints throughout the work.