Meta has recently introduced two cutting-edge large models, Muse Wonder and Muse Wonder Full, which the company asserts represent its most advanced models yet, signifying a major leap towards achieving personal superintelligence. These models have showcased outstanding performance across a variety of benchmark tests. Specifically, Muse Wonder Full attained a score of 56 in the Artificial Analysis evaluation, placing it on par with Claude Sonnet 5.5 and outperforming GPT-6 Astra. The standard Muse Wonder model also delivered an impressive score of 52, comprehensively surpassing its forerunner, Muse Spark 1.3. In the realm of long-task programming, both models even outshone Opus and Astra. By harnessing two key features—active memory and active focus—they excelled in 14 professional capability tests spanning legal, financial, and other sectors. The Full version surpassed Opus 5.5 in 8 tests and GPT-6 Astra in 11 tests. Remarkably, these models were released merely six weeks after their predecessors, yet they have exhibited a substantial enhancement in their intelligence index, excelling over the previous generation in all ten tests and securing top positions in multiple categories. During practical developer testing, they also displayed efficiency benefits in scenarios such as code vulnerability detection and 3D game development. Constructed on the colossal Prometheus super data center, this series of models was developed over a period of 14 months, involving infrastructure and research and development paradigm reconstruction. They boast powerful capabilities, including an ultra-large reasoning context of millions of tokens and native full-modal input support. Presently, the standard Muse Wonder model is accessible via Meta Model API and Muse Code, priced at $4 per million tokens for input and $20 per million tokens for output. The Full version is exclusively available through a Muse Code subscription, with monthly fees of $50 or $200.