Meta launches Muse Wonder models to strengthen AI capabilities

Meta has unveiled two new artificial intelligence models, Muse Wonder and Muse Wonder Full, designed to tackle complex tasks, write code and retain information over extended periods, as the company intensifies its rivalry with the leading AI developers.

According to a TechCrunch report, Meta’s Superintelligence Labs released the models on Monday, 12 October, signalling a shift towards more capable systems able to handle demanding, time-consuming work. They build on the company’s earlier Muse Spark model, which prioritised speed and rapid responses. Furthermore, Meta’s model cards indicate that the new systems perform comparably to GPT-6 Astra and Fable 5.1 on several benchmarks, although independent evaluations are still needed to verify those claims.

A defining feature is what Meta calls “active memory”, which enables AI assistants to retain and reuse information from earlier interactions over days, weeks or even months. Research scientist Kavya Kopparapu said the team had invested considerable effort in ensuring that the models remember what matters to users. Moreover, the company claims that Muse Wonder outperformed Fable, Opus and Astra on MemoryBench, a test designed to gauge long-term learning and information retention.

Read More: Common mistakes that can get your WhatsApp account banned

The two models differ in availability and pricing. Muse Wonder costs $4 per million input tokens and $20 per million output tokens through Meta’s model API, whereas Muse Wonder Full is restricted to Muse Code subscribers and is not offered through the API. Developed at Meta’s Prometheus data centre in Ohio, the models represent a significant step in the company’s drive to strengthen its advanced AI portfolio and compete more directly with OpenAI and Anthropic.