Meta releases Muse Wonder models for coding and long tasks
Meta has released Muse Wonder through its model API and coding service, while Muse Wonder Full is available to coding-service subscribers. The company says the models are designed to retain task context and keep working through complex steps. Independent coverage confirms the release but does not establish that Meta’s benchmark claims hold across real-world tasks. A later rollout to the Muse personal assistant is planned; this announcement concerns the underlying models.
Artificial Intelligence··Evening
Two Wonder models enter developer services
Meta released Muse Wonder and Muse Wonder Full on October 12 as new models for sustained AI tasks. Muse Wonder is available through the Meta Model API and Muse Code, the company’s coding service. The larger Full version is available to Muse Code subscribers. Independent reporting also identifies this as a model release. The personal Muse assistant had already appeared in apps; the new event concerns the underlying model family and its access channels.[1], [2]
Meta says the models are meant to hold task information across a long conversation and continue work when a first approach fails. It calls those behaviours Active Memory and Active Focus. In coding, the intended workflow involves tracking requirements, editing files, testing changes and revising the result. Those are the capabilities the company designed and demonstrated; they do not establish that every user task can be completed reliably.[1], [2]
Long-task claims rest on company evaluations
The announcement uses software-project and focus benchmarks to describe progress during extended work. It also cites an evaluation of legal tasks involving documents and instructions. Meta presents these as performance evidence from its own testing. Its examples include asking users for clarification and checking with them before important actions. Such demonstrations show intended behaviour, while the extent of reliability outside those tests remains to be established.[1]
TechCrunch describes Wonder as a larger, more capable successor to the lightweight Muse Spark. Its account confirms that Full is confined to the coding subscription rather than the general API. That distinction matters because the two releases do not have identical access routes. Meta has separately said the models will later reach the Muse assistant, initially for subscribers. The planned app deployment should not be confused with today’s developer availability.[1], [2]
Safety details accompany the rollout
Meta says its release process addresses personal information, adversarial instructions and misuse in sensitive chemical, biological, radiological, nuclear and cyber areas. The company describes safeguards but says a fuller report with additional evaluations will come later. The present announcement therefore gives the scope of its safety work without supplying all of the promised results. It also does not announce an open release of the models’ weights.[1]
The two models were trained at Meta’s Prometheus data centre, according to the company. Meta argues that a stronger base model can carry skills into tasks it was not specifically trained to solve afterward. That remains a company account of why the new release might work across varied jobs. The concrete change for developers is access to Wonder through the API and coding service, with Full under a narrower subscription and assistant rollout still ahead.[1]
