Meta Puts Two New Models Behind Its Personal AI Push
Meta has introduced Muse Wonder and Muse Wonder Full, two models aimed at the kind of work that runs longer than a single prompt.
- Models: Muse Wonder and Muse Wonder Full
- Availability: the Meta Model API and Muse Code, with the Full model for Muse Code subscribers
- Built on: the Muse Spark model, with advances in pretraining and reinforcement learning
- Trained at: Meta's Prometheus data centre
- Target tasks: agents, memory, coding, and domain work in law, finance and health
What Muse Wonder Is Built To Do Differently
Two features carry the pitch. Active Memory keeps relevant context across conversations and retrieves it when needed. Active Focus is aimed at goals rather than answers: trying alternative approaches, working around obstacles, adapting to new information, and checking with the user before an important action.
Both describe the same shift. The useful unit of work stops being a reply and becomes a task that persists, which is also where most assistant deployments currently fall over.
The Coding Case And The Generalisation Claim
In software work the models are built to manage requirements, make changes across a codebase and run repeated testing and revision, trained on long-running tasks so progress holds from small applications through to complex software.
Meta also evaluated them on unfamiliar tasks, and states that Muse Wonder Full performed better there than the Muse Spark 1.3 performance trend would predict, attributing the difference to stronger underlying models. That is the claim worth independent testing, because generalisation beyond the training target is precisely what nobody can yet take on trust.
Why The Safety Section Matters Commercially
A personal agent that can read messages, calendars and private documents is a different product from a chatbot, and a different liability. Meta states the models were trained to resist adversarial instructions and handle private information carefully, and claims its strongest safeguards so far against chemical, biological, radiological and nuclear misuse and against cyberattacks.
For marketers the practical read is narrower. A model that holds context across a campaign and will stop to ask before acting is a far easier thing to put near live budgets than one that forgets between sessions.
More ad tech coverage on Ad Tribe.