Meta Unveils Muse Wonder AI Models, Says It Has Reached the Frontier as Its Stock Jumps 4%

Meta Platforms (META) unveiled a new series of AI models on Monday, Oct. 12, called Muse Wonder and Muse Wonder Full, which it describes as its most capable models to date. Meta shares rose about 3.8% to $745.90 at midday. The company said the models are built for personal and professional tasks, and CEO Mark Zuckerberg told the Financial Times that Meta has reached the AI frontier and caught up with competitors.

What Meta launched

Meta said Muse Wonder will soon power its Muse assistant, starting with subscribers. Index scores are as reported by the New York Times. Meta has not said whether it will release the model weights, according to TokenPost.

Active Memory and Active Focus

Meta highlighted two features. Active Memory lets the models retain relevant personal context across conversations and bring it back up when it is useful for a later task. Active Focus is designed to keep a model working on complex, multi-step jobs: it tries other approaches when it hits an obstacle, recovers from errors, asks for clarification when needed and checks in before doing anything important.

Meta pointed to law, finance and health, along with coding, as target areas. Examples cited in coverage include navigating insurance claims, interpreting leases and building applications.

How it compares

On the Artificial Analysis Intelligence Index, Muse Wonder Full scored 56, second to Anthropic's Claude Opus 5.5 at 58. The standard Muse Wonder scored 52. In the comparisons Meta supplied, Wonder Full leads on legal-agent tasks, business workflow automation and long-horizon software engineering, though competitors lead some other benchmarks.

Treat the results with some caution. One outlet reported a separate Meta-reported measure, the "Personal Superintelligence Index," on which Wonder Full scored 59.1 against 55.2 for Claude Opus 5.5 and 54.5 for OpenAI's GPT-6 Astra, but it was not clear who ran that test. Meta's claims have not yet been independently verified across the board.

Infrastructure and safety

The models were trained at Meta's gigawatt-scale Prometheus data center using expanded pretraining and reinforcement learning. Meta said they perform better on unfamiliar tasks than its earlier Muse Spark 1.3. On safety, Meta says the models are trained to avoid sharing private information without permission and to decline requests to help with harmful acts, with stronger protections against unauthorized actions, privacy violations and cyber misuse. A full safety evaluation report is "coming soon."

Zuckerberg also told the Financial Times that Meta is "already working on the next step up," rejected calls to slow AI development, and said responsibility for model safety should fall on individual labs.