After more than a year of work and tens of billions of dollars of investment to restructure artificial intelligence studies, Meta introduced two new artificial intelligence models named Muse Wonder and Muse Wonder Full. The company claims that with these models, it has reached the performance level of the most advanced systems in the field of artificial intelligence. According to Meta’s statement, the new models contain significant progress towards the goal defined as personal superintelligence. The models, which offer advanced capabilities especially in law, finance and health-related tasks, also promise high-level performance in the field of software development. In addition, two new features called active memory and active focus aim to enable artificial intelligence systems to work with users on longer and more comprehensive tasks.
According to the information shared by Meta, Muse Wonder and Muse Wonder Full are not designed as artificial intelligence models that only answer questions. The company wants these systems to be able to track tasks consisting of different stages and help users achieve their goals. The active memory feature developed for this purpose allows the information shared in previous conversations to be reused when necessary. Active focus, on the other hand, aims to prevent difficulties encountered during a complex task from interrupting the work. Thus, the aim is for the models to go beyond one-time commands and perform longer-term operations.
Meta added active memory and focus features to Muse Wonder models
Active memory, one of the striking features of the new models, focuses on preserving the context between different conversations with the user. Meta explains this feature as carrying information that is valuable to the user between conversations and remembering details about the current task when necessary. Thus, it is aimed to reduce the need for users to re-share the information they have previously provided in each new conversation. For example, during a long-term study, it may be possible to reuse preferences or requirements determined in previous interviews. However, Meta has not yet shared details about what information this feature will store and for how long, and what control users will have over memory.
The active focus feature aims to enable artificial intelligence agents to continue working in the face of challenging tasks. According to Meta’s statement, this system will be able to continue progressing towards the set goal when the task becomes complex or unexpected difficulties are encountered. Especially in software development, processes that require extensive research and multiple stages, such a capability can make the job easier for users. On the other hand, the ability to complete long-term tasks reliably depends on the model’s ability not only to perform operations but also to identify and correct errors when necessary. How Meta’s new models will perform in such scenarios can be understood more clearly with real usage experiences.
The company also plans to integrate Muse Wonder models into the personal artificial intelligence agent Muse. It is stated that this integration will be offered to paid subscribers in the first stage. Being able to use active memory and active focus features through Muse can make it possible for the artificial intelligence assistant to follow ongoing work in different sessions. In addition, it is aimed that users will be able to manage tasks that cannot be completed with a single command with less intervention. Meta’s statements show that new models are planned to be evaluated in artificial intelligence agents that perform direct transactions, as well as in chat-based use.
Meta’s new models also attract attention with the high budget the company allocates for artificial intelligence infrastructure. It is stated that the company may spend up to 145 billion dollars this year to support artificial intelligence studies. Speaking to the Financial Times, Meta CEO Mark Zuckerberg said that the level reached requires a high amount of computing capacity and investment. Zuckerberg also stated that the meticulous scientific studies carried out by the company’s super intelligence team are also decisive. Meta’s restructuring of its artificial intelligence organization over the last year reveals that the development process of new models is not just about hardware investments.
Muse Wonder is available for developers via Meta’s API service and within Muse Code. The pricing of the model was announced as $4 for one million entry tokens and $20 for one million exit tokens. This pricing requires consideration of the total cost of use, especially for developers who process high volumes of transactions. The more advanced version, Muse Wonder Full, is offered to users who have one of Muse Code’s $50 monthly Power Usage or $200 Maximum Usage subscriptions. Thus, Meta imposes a usage-based fee via the API for the standard model, while limiting access to Wonder Full to certain subscription packages.
The launch of Muse Wonder and Muse Wonder Full is one of the concrete results of Meta’s work on developing high-capacity artificial intelligence models. Active memory and active focus features can make artificial intelligence agents more functional, especially in long-term tasks. However, the company’s claim that it has reached the same level as the most advanced artificial intelligence systems needs to be evaluated with independent performance tests and real usage results. It is also necessary to examine talents in sensitive fields such as health, law and finance in terms of accuracy, reliability and data privacy. The features and pricing announced by Meta show that the new models are initially aimed more at developers and heavy artificial intelligence users.