Article · AI translation
IT Home reported on November 7 that Meta released its Muse Wonder series of models last month. On November 6 local time, Meta announced dedicated versions of its existing models trained on Muse Wonder and began rolling them out to all Muse and Meta AI users.
When it released Muse Wonder, Meta introduced 2 core features for the first time: Active Memory and Active Focus. Both capabilities have now been made available across the dedicated models in Muse and Meta AI.
According to Meta, Active Memory enables Muse and Meta AI to remember key information accurately and retrieve it when earlier conversations are relevant to the current situation, making it easier to build on past conversations.
The upgraded models in Muse are 2.4 times less likely to forget key information needed to achieve a goal.
IT Home learned that Active Focus helps Muse stay on course toward a goal even when unexpected challenges arise. In Muse, the new models' task completion rate increased by 67% when routine tasks took an unexpected turn.
Further reading
Muse Wonder is a new generation of AI models that Meta officially released on October 12, 2026 (U.S. Eastern Time). The series includes Muse Wonder, designed for fast, frontier-level performance, and the full version, Muse Wonder Full. Zuckerberg said it is larger and more capable than the earlier Muse Spark and represents an important step toward Meta's goal of “personal superintelligence.”
Muse, mentioned in this article, is a personal AI agent Meta launched in September 2026. It can help users with everyday tasks such as buying groceries and booking trips, and connects to Meta's social apps. After launch, it topped the download charts in the U.S. Apple and Android app stores. Meta AI is Meta's AI assistant for a broader audience.
Regarding the 2 core features, Meta describes “Active Memory” as a model's ability to retain important background information about a user across conversations and proactively retrieve details relevant to the current task, while persistent storage, deletion, review and authorization are controlled by the memory module of the application above it. “Active Focus” means the model tries different approaches to challenging goals, overcomes obstacles and adjusts its actions based on new information.
According to internal evaluations released by Meta, after the new models were integrated, Muse correctly used key context in 90% of tasks, up from 76%. The Muse models scored 47% on Meta's “Personal Superintelligence Index,” up from 38%. Meta reported all these figures itself; no independent third-party evaluation is currently available.
As for access, Muse Wonder is available to developers through the Meta Model API and Muse Code, while Muse Wonder Full is available only to Muse Code subscribers. The new models in the personal AI assistant Muse are being made available to subscribers first, with subscriptions priced at $20 or $100 per month. The Muse Wonder series is also closed-source.
Publicly available information shows that Muse Wonder was trained on Meta's gigawatt-scale Prometheus supercomputing cluster and natively supports a context of up to 1 million tokens, as well as text, image, video and audio input. The series is also the second-generation result released about 14 months after Meta Superintelligence Labs (MSL) rebuilt its technology stack.
Related reading:
“Meta Releases Flagship Muse Wonder AI Model Series to Compete With Anthropic and OpenAI”
“Meta Releases Muse Wonder Model; Zuckerberg Says Company Has Caught Up With Rivals at the AI Frontier”