Meta Unveils 'Muse Wonder': The Tech Giant's First Frontier AI Model Unleashed
Meta's Superintelligence Lab has unveiled Muse Wonder and Muse Wonder Full, its first frontier models designed for enhanced capabilities in complex tasks like coding and comparable to leading AI models. These models feature "active memory" for long-term information retention and will power the Muse agent, representing a major step towards personal superintelligence.
Meta’s Superintelligence Lab has officially unveiled its inaugural frontier models, Muse Wonder and Muse Wonder Full, marking a significant stride in the company’s artificial intelligence development. These new models represent a departure from the earlier Muse Spark, which was designed for lightweight operations and low latency. In contrast, Muse Wonder and Muse Wonder Full are substantially larger and engineered for superior capability, demonstrating higher benchmark scores and enhanced performance across intricate tasks, including complex coding challenges.
Internal testing conducted by Meta, as detailed in the models' card, reveals that Muse Wonder and Muse Wonder Full achieve benchmark scores comparable to industry leaders such as Fable 5.1 and GPT-6 Astra. Meta’s official blog post emphasizes that these models signify “the next step on our scaling ladder and a major leap toward personal superintelligence,” underscoring their advanced nature and ambitious goals. The increased scale and capability are also reflected in their pricing structure.
Access to Muse Wonder is facilitated through the Meta Model API, with costs set at $4 per million input tokens and $20 per million output tokens. This pricing positions Muse Wonder as a direct competitor in cost to other high-tier models like GPT-6 Sol or Opus 5.5. Notably, Muse Wonder Full will not be available via the API; instead, its access will be exclusively granted to Muse Code subscribers. Both new models are slated to play a crucial role in powering the Muse agent, further integrating them into Meta's AI ecosystem.
A key innovation highlighted by Meta is the new models' sophisticated ability to retain and synthesize user-provided information over extended periods, a feature the company terms “active memory.” This capability aligns with Meta’s broader objective of developing personal assistants adept at performing tasks continuously. To validate and improve this crucial recall function, Meta specifically emphasized its performance on the MemoryBench benchmark, a rigorous test designed to evaluate continual learning and memory construction frameworks. According to Meta’s assessments, Muse Wonder now surpasses Fable, Opus, and Astra in MemoryBench scores, signifying a significant advancement in long-term memory. Research scientist Kavya Kopparapu reiterated the company's commitment, stating in a launch video, “We put a tremendous amount of effort into making sure your model can remember what’s important to you over days, weeks, months, and beyond.”
The release of Muse Wonder and Muse Wonder Full marks the first frontier model launch since Scale AI founder Alexandr Wang undertook a significant restructuring of Meta’s AI initiatives in August of the previous year. Prior to this, Meta had focused on lighter models like Muse Spark and the Muse agent, largely avoiding the state-of-the-art model releases characteristic of companies like OpenAI and Anthropic. The development and training of these groundbreaking models took place at Meta’s Prometheus data center, located in central Ohio, which became operational earlier this year.