Superintelligence Lab released two new advanced models, Muse Wonder and Muse Wonder Full, on Monday, marking the company's first attempt at launching cutting-edge models.

Meta The previously launched Muse Spark was intended to be a low-latency, lightweight model, while the Wonder series of models are larger in scale. Meta claims to have stronger capabilities, achieving higher scores in benchmark tests and performing better on complex tasks such as coding as well.

According to the test results displayed in the model card by Meta, the benchmark scores of Wonder and Wonder Full are comparable to those of Fable 5.1, as well as GPT-6 and Astra.

Meta writes in a blog post about these models: 'They collectively mark the next step on our roadmap for expansion, and also represent a significant leap towards individual superintelligence.'

This scale is also reflected in pricing: when accessing via Meta Model API, the price for every million inputs of token is $4.00, and for every million outputs of token it is $20.00. The cost is roughly equivalent to GPT-6 Sol or Opus at $5.5.

It is worth noting that Wonder Full will only be available to subscribers of Muse Code, and will not be provided through API. Both models will also be used to drive Muse and agent.

Meta has always been focused on developing personal assistants that can perform tasks over the long term, therefore the company places particular emphasis on the new model's ability to retain and integrate information provided by users. Meta refers to this feature as “active memory”.

In order to enhance memory capabilities, Meta emphasized the MemoryBench benchmark, which tests the framework for continuous learning and memory building. According to the company's tests, Muse Wonder has scored higher on MemoryBench than Fable, Opus, or Astra.

Research scientist Kavya Kopparapu who participated in the video release stated: "We have invested a great deal of effort to ensure that your model is able to remember the things that are important to you for several days, weeks, months, and even longer."

These new models are the first cutting-edge models released by Meta since Scale AI founder Alexandr Wang restructured the Meta AI business last August. Since then, Meta has released lightweight Muse Spark models and Muse agent, but has been avoiding releasing such advanced models as OpenAI and Anthropic.

According to the article, the relevant model is trained at the Prometheus data center located in central Ohio, where Meta is situated, and this data center was put into operation earlier this year.