Meta Superintelligence Lab released Muse Wonder and Muse Wonder Full on Monday, marking the company’s first release of frontier AI models. Meta emphasizes their ability to work with user information over longer periods. Its launch description also credits them with better results on complex tasks, including coding, than the lightweight Muse Spark.
Active memory is designed to retain and connect information
Meta calls the models’ ability to retain and connect user-provided information “active memory.” At the launch, research scientist Kavya Kopparapu explained that the team had worked to help the models remember what matters to users across days, weeks, months and longer.
The company highlights MemoryBench for this area, a benchmark that tests continual-learning and memory-building systems. According to Meta’s testing, Muse Wonder scores higher on it than Fable, Opus or Astra.
Larger than Spark, with scores comparable to Fable 5.1
Muse Spark was designed as a lightweight model focused on low latency. Both Wonder models are larger, and Meta’s launch description credits them with higher benchmark scores and better performance on complex tasks, including coding.
According to Meta, tests published in the model card show Wonder and Wonder Full achieving benchmark scores comparable to Fable 5.1 and GPT-6 Astra.
Wonder through the API, Wonder Full with Muse Code
Access to Muse Wonder through the Meta Model API will cost $4.00 per million input tokens and $20.00 per million output tokens. Muse Wonder Full will be available only to Muse Code subscribers and will not be offered through the API.
Both new models will also power the Muse agent, which Meta introduced before this release.
First frontier release since the AI reorganization
The release follows the restructuring of Meta’s AI efforts by Scale AI founder Alexandr Wang in August of last year. Since then, the company has released Muse Spark and the Muse agent; Wonder and Wonder Full mark its first frontier-model release since that reorganization.
According to Meta’s post, the new models were trained at its Prometheus data center in central Ohio. The facility began operating earlier this year.
