Meta released two AI models on 12 October, Muse Wonder and Muse Wonder Full, and says both lead the field at writing software and at carrying out tasks for a user. The announcement came from Meta AI Research in a post titled “Muse Wonder: A Frontier Model For Personal & Professional AI Agents.” Its centerpiece is a memory feature Meta calls Active Memory. Every performance claim in the post is Meta’s own.
Start with the “frontier” label, because it is easy to misattribute. The post calls the pair Meta’s “most capable models” and describes Muse Wonder as a frontier model, so the frontier claim is the company’s, not an outside verdict. It does not call them Meta’s first frontier-class models. What it does say is that they are “the next step on our scaling ladder” after Muse Spark, an earlier model in the same family, and that Meta has spent fourteen months rebuilding its AI effort from the ground up.
Active Memory has a specific definition in the post: “carrying important context about you across conversations and actively retrieving the details relevant to the task at hand.” In plain words, the assistant keeps hold of facts about you from one chat to the next and pulls up the relevant ones when a job calls for them, instead of starting blank every time. Meta pairs it with Active Focus, meaning the model keeps working on a hard goal, tries other approaches when it hits a blocker, asks for clarification, and checks in before important actions.
The post does not say what the memory stores, for how long, or whether users can see and delete it. That gap matters because Meta itself notes that a personal agent may handle messages, calendars, and private documents.
On the two models, the post is vague. Muse Wonder is available now through the Meta Model API and Muse Code, Meta’s coding product. Muse Wonder Full is available now only to Muse Code subscribers. Meta says Muse Wonder will later power features inside its Muse assistant, rolling out first to subscribers. Nothing in the post states how the two differ in size, price, or ability, beyond both resting on “larger, stronger base models.” Any description of Full as the bigger or smarter one would be a guess.
The benchmark claims need the same care. Meta says the models reach “state-of-the-art results” for agents, memory, and coding, compared against other frontier models. The post names tests including Harvey’s Legal Agent Benchmark, FocusBench, and SWE-Marathon, which covers long-running software projects. It names no independent evaluator, and the text of the post carries no scores; the comparison tables sit in graphics and a separate report. A fuller safety report is promised “soon.”
One experiment is worth a note. Meta plotted Muse Wonder Full against Muse Spark 1.3 on tasks the models were and were not trained to do, and says the newer model beats the older trend on unfamiliar tasks. Meta frames this as a response to complaints about models tuned to ace benchmarks. It is also Meta grading its own homework, with Meta choosing the tasks.
The larger context is that a model’s memory is hard to judge from a leaderboard. A coding score can be checked by anyone with the same test. Whether an assistant recalls the right detail about you three weeks later can only be seen in use. Meta is also putting the stronger version behind a subscription, which signals where it expects the money to come from.
For developers on the Meta Model API, the near-term job is to test memory across separate sessions on your own data before trusting the headline scores, since Wonder Full is not on offer there.
Based on Meta AI Research’s blog post of 12 October 2026.
