Meta launches Muse Wonder model with high benchmark score

Meta has launched the large language model Muse Wonder, which is larger than Muse Spark, successfully bringing the company back into a leading position after resetting its AI development process for 14 months. The top version, Muse Wonder Full, achieved a benchmark score of 56 on the Artificial Analysis Intelligence Index, ranking second overall and only behind Claude Opus 5.5, which scored 58. Muse Wonder outperformed all other models, including Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon. The standard version of Muse Wonder received a similar benchmark score, but the Full version is designed to handle a wider range of unfamiliar tasks.

Meta stated that Muse Wonder is designed for agentic work, referred to as Personal Superintelligence, suitable for both personal and professional use. It emphasizes four key features: Domain Expertise, which provides knowledge in specialized fields such as law, finance, and health, where the model scored well in benchmarks; Active Memory, a memory feature that retains important information from past chats for use with current prompts; Active Focus, which directs the model to "try" to achieve goals using new methods; and Frontier Coding, which enhances coding capabilities, outperforming Claude Opus 5.5.

Muse Wonder is now available through the Meta Model API and Muse Code, while the Muse Wonder Full version will be available exclusively through Muse Code soon. Additionally, the Muse Wonder model will be utilized with the Muse Agent assistant service. Blognone reported the story.