Meta has launched Muse Wonder and Muse Wonder Full, two artificial-intelligence models it describes as its most capable to date and a step toward CEO Mark Zuckerberg’s goal of “personal superintelligence.”
The models are designed for personal and professional agents handling tasks such as insurance claims, leases, financial information, coding and computer-based workflows. Meta says Active Memory lets them retain important information across conversations, while Active Focus helps them continue difficult work, try alternative approaches, recover from errors and request confirmation before taking significant action.
Muse Wonder is available through Meta’s developer API and Muse Code, with a one-million-token context window. Muse Wonder Full is not offered through the API; access is limited to Muse Code’s Power Usage plan at $50 per month and Maximum Usage plan at $200 per month.
Meta’s published evaluation table lists an Artificial Analysis Index score of 56 for Wonder Full and 52 for Wonder at maximum reasoning effort. Meta also says Wonder performed better than those competing models on its MemoryBench testing, although those comparisons remain company-reported results. AI CEO Bindu Reddy disputed the performance claims and said her organization would publish detailed benchmark results.
The launch marks a shift for Meta’s AI strategy from distributing lightweight or open-weight systems toward competing directly with leading proprietary models. The company said both Wonder models will power its Muse agent, with broader Muse integration planned to begin with subscribers. Meta has not said whether either Wonder model will be open-sourced, while it has separately released the weights for Muse Glimmer.
Meta said it trained the models at its Prometheus data center in central Ohio and built safeguards intended to prevent unauthorized sharing of private information, reject harmful requests and support defensive cybersecurity work. The company said further safety evaluations are forthcoming, leaving independent testing and the models’ eventual reach across Meta’s consumer products as the next measures of whether the launch’s performance claims translate into wider use.