Meta unveils Muse Wonder and Muse Wonder Full, agentic AI models with 1M-token context

Meta on Monday launched two new “frontier” artificial intelligence models for software agents, Muse Wonder and Muse Wonder Full, and made them available immediately through its developer tools. In a newsroom post announcing the release, Meta called them “our most capable models to date” and said they were built for “personal and professional agentic tasks,” meaning systems that can carry out multistep work on a user’s behalf.

Access is rolling out in stages. Muse Wonder is available now to developers through the Meta Model API and Muse Code, the company’s coding product. Muse Wonder Full is available now to Muse Code subscribers. Meta also said Muse Wonder will soon begin powering experiences in the Muse consumer app, with the rollout starting with subscribers.

Meta is pitching the models around two main features it calls Active Memory and Active Focus. In a research blog post, the company said Muse Wonder “brings together Active Memory and Active Focus to keep track of what matters to you and work diligently toward what you want to accomplish.” Meta says Active Memory lets the system carry relevant context across conversations, while Active Focus is designed for sustained problem-solving, including trying alternate approaches, recovering from errors and checking in before taking important actions. According to the Muse Wonder model card, the models accept text, image, video and audio inputs and generate text output. The same document lists a context window of 1 million tokens at inference, a measure of how much information the model can take into account at once.

The launch marks a more aggressive push by Meta into top-tier AI systems, and a more controlled one. Meta says the models were trained at its gigawatt-scale Prometheus data center and presents them as direct competitors to other frontier systems. In the Muse Wonder model card, Meta compares Wonder and Wonder Full against named rivals including OpenAI-branded GPT-6 Astra and Anthropic-branded Claude Fable 5.1 and Claude Opus 5.5 across categories such as agent tasks, coding and memory. Those comparisons come from Meta’s own materials and were not independently verified.

Meta’s release also included a concise but notable safety disclosure. Under the company’s Superintelligence Scaling Framework, the model card says unmitigated Muse Wonder Full could reach a “high risk threshold” in the chemical and biological, or CBRN, domain. Meta said it applied safety post-training and system-level classifiers — filtering and monitoring systems layered around the model — and that with those mitigations the deployed version is at “moderate or lower” risk across domains, which the company said supports deployment. Meta said a fuller report with additional safety evaluations will be published soon.

The release stands in contrast to Meta’s earlier approach this year. In 2026, the company released Muse Glimmer, a 30 billion-parameter open-weight agentic model under the Apache 2.0 license, allowing outside developers to download and run it more freely. Muse Wonder and Wonder Full are not being released that way, at least for now. Meta’s press materials describe a cloud-served launch through its own products rather than an open-weight release.

That shift also ties the new models more closely to Meta’s broader consumer assistant push. The company launched the Muse app and related consumer integrations in September, and now says Wonder will soon power those experiences. Meta also updated its Superintelligence Scaling Framework on Oct. 2 and said Wonder was trained, evaluated and released under that framework.