The models are Meta’s most capable AI systems to date, according to the company, and are designed to power personal and professional AI agents capable of carrying out extended tasks with limited human intervention.

The launch marks another step in Meta’s push to develop AI systems that can do more than respond to individual prompts, instead working toward users’ goals over time, adapting to changing circumstances and managing complicated workflows.

Among their key capabilities is Active Memory, a feature that allows the models to retain important information about users across multiple conversations and retrieve it when relevant.

Meta has also introduced Active Focus, which is designed to help AI agents continue working when they encounter obstacles by exploring alternative approaches, adapting to new information and requesting clarification when necessary.

The models can also delegate tasks to smaller AI agents, recover from errors without abandoning their original objectives and work in specialized fields such as law, finance and health, according to Meta.

The company said the systems could assist with a range of everyday and professional activities, including reviewing lease agreements, navigating insurance claims, analyzing personal finances and developing software applications.

In a move to make the technology available to developers, Meta has released Muse Wonder through its Meta Model API and Muse Code programming tool.

Muse Wonder Full, the more capable of the two models, is available through paid Muse Code subscriptions, with plans starting at $50 per month for access to Meta’s most advanced models.

Meta plans to integrate Muse Wonder into its Muse personal assistant at a later stage, initially making the model available to subscribers.

To support its claims about the models’ capabilities, Meta published benchmark results that it said showed the systems matching or outperforming selected models from OpenAI and Anthropic in coding, memory and professional tasks.

However, the results represent the company’s own assessments, and the models’ broader performance will require further evaluation through independent testing.

Alongside the performance claims, Meta said it had trained the Wonder family to protect private information, refuse assistance with harmful activities and strengthen defenses against cyberattacks.

The company said it would publish a more detailed safety report at a later date.

The launch is part of Meta’s broader strategy to develop what CEO Mark Zuckerberg has described as “personal superintelligence” — AI systems capable of understanding users’ goals, preferences and long-term needs.

The approach reflects the company’s ambition to move beyond general-purpose chatbots toward AI assistants that can retain context, make progress on complex tasks and provide more personalized support.

Unlike some of Meta’s earlier AI models, which were released with downloadable weights, the company did not announce plans to make the Muse Wonder models open source.