Muse Wonder is betting on memory, but it was tested by Meta itself

Meta released Muse Wonder and Muse Wonder Full and immediately put its weight behind them: both models are positioned as larger and more capable than Muse Spark, especially for complex tasks such as programming. According to the company’s internal tests, their results are comparable to Fable 5.1 and GPT-6 Astra.

Meta is placing particular emphasis on “active memory.” The models are supposed to retain and connect information about the user over long periods of time. In MemoryBench, which specifically tests memory and continual learning, Muse Wonder scored higher than Fable, Opus, and Astra, according to Meta’s own data.

But there is an important point for those who are actually choosing a model. Through the Meta Model API, Muse Wonder costs $4 per million input tokens and $20 per million output tokens. Muse Wonder Full is not available through the API; it can only be used by Muse Code subscribers. Both models will also work in the Muse agent.

The comparisons and MemoryBench results were provided by Meta itself. There is no independent verification of the memory claims in the article, nor are there details about how well it works in everyday tasks. And for those who specifically need Wonder Full, access is already limited to subscribers.

For which tasks would it be most useful for you if AI could remember context across different days? Source: AI News & Artificial Intelligence | TechCrunch