Blog
Meta launches Muse Wonder and Muse Wonder Full, models built for personal and work agents
On 12 October 2026 Meta introduced Muse Wonder and Muse Wonder Full, its most capable models so far, aimed at personal and professional agent tasks. Muse Wonder is available through the Meta Model API and Muse Code; Muse Wonder Full is for Muse Code subscribers. A full safety report is still to come.
On 12 October 2026 Meta introduced the Muse Wonder series: two models, Muse Wonder and Muse Wonder Full, which the company calls its most capable so far. Meta is pitching them at one job in particular: agents that carry out personal and professional tasks for people, not just answer questions.
What Meta released
According to Meta's announcement:
- Muse Wonder is available now through the Meta Model API and Muse Code.
- Muse Wonder Full is available now to Muse Code subscribers.
- Muse Wonder will "soon" power experiences in Meta's Muse app, rolling out to subscribers first.
- Meta also plans to use Muse Wonder to build smaller specialised models for everyone using Muse.
Meta says both models were trained at its gigawatt-scale Prometheus data center and build on the scaling approach it described when it launched Muse Spark.
The two ideas Meta is selling: Active Memory and Active Focus
Meta groups the models' agent abilities under two labels.
Active Memory
This is Meta's term for keeping important context about a user across conversations and pulling up the right details when they become relevant to a task. Meta argues that reliable memory is one of the most important requirements for a useful personal agent.
Active Focus
This describes how the model keeps working on a hard goal: trying other approaches, working around blockers, and adapting to new information. Meta says the model also asks for clarification when needed and checks in before taking important actions. It measures this with benchmarks it calls FocusBench and SWE-Marathon, the second covering long-running software projects.
Coding and domain work
Meta says both models deliver "frontier coding performance" and were trained on long-running tasks so they can keep making progress across repeated test-and-fix cycles. It also highlights work in law, finance and health, citing Harvey's Legal Agent Benchmark as one of its evaluations.
Demos on the announcement page include building a playable 3D kitchen game from images, planning a running schedule in a browser, preparing a wedding seating chart, and designing and ordering custom merchandise from photos on an Instagram profile.
Meta says the models lead on key agent, memory and coding benchmarks. It publishes comparison charts and points to a separate evaluation report. Those are Meta's own claims, and independent testing will matter more than launch charts.
Generalisation, not just benchmarks
Meta spends part of the post on the "benchmaxxing" complaint, where a model scores well on popular tests but struggles with unfamiliar work. It compares Muse Spark 1.3 with Muse Wonder Full and says the newer model does better on tasks it was not specifically trained for than the older model's trend would predict. Meta credits the larger, stronger base models underneath.
Safety
Because personal agents may touch messages, calendars and private documents, Meta says it trained the Muse Wonder family to resist adversarial instructions and handle private information carefully. It says Muse Wonder has its strongest safeguards yet against chemical, biological, radiological, nuclear and cyber misuse, while keeping advanced cyber abilities available for defensive use under its Meta Superintelligence Scaling Framework.
Meta calls Muse Wonder its "most aligned model to date". But the full safety report is not out yet: Meta says it "will be available soon".
What this means for you
If you already use the Meta Model API or Muse Code, you can try Muse Wonder now. For anyone thinking of handing an agent their inbox, calendar or accounts, Meta's emphasis on checking in before important actions is a useful design point. It is still worth waiting for the full safety report and independent results before trusting it with anything sensitive.