Muse Wonder puts meta at the edge of artificial intelligence
The biggest meta model competes with the best models in the world and has new features to prevent agents from leaving.
When Mark Zuckerberg relaunched his company's AI efforts last year, it seemed like a very expensive gamble. Meta's latest AI models, the Muse Wonder and Muse Wonder Full, seem to show that this bet is finally paying off by putting the company back on the cutting edge of AI development.
Meta says Muse Wonder Full has borderline capabilities, meaning it's one of the most capable AI models in the world. "Together, they represent the next step in our scalability ladder and a giant leap toward personal superintelligence," the company said in a blog post today.
Muse Wonder models will be available via API and through Muse Code, a tool for writing artificial intelligence software. Meta says that the best version of this model will enhance capabilities across all of its platforms, including the Muse personal agent.
Zuckerberg revamped his company's AI research last year by sinking billions of dollars into the recently formed Meta Superintelligence Labs and hiring former Scale AI CEO Alexander Wang to lead it. This effort has sometimes been difficult, and some researchers have not stuck around for long.
Meta was once one of the pioneers of artificial intelligence. Its Llama series models were the first frontier open weight offerings from a major lab. Llama, which could be downloaded and run on personal hardware, was widely used by researchers and startups.
But the company fell behind after delivering several disappointing versions of the model, as Anthropic, OpenAI, and Google quickly introduced more capable models.
Meta introduced the first model of its new lab, the Muse Spark, in April of this year. The company has also released a small open-weight model, the Muse Glimmer, and says it plans to release larger open-weight models in the future. It did not reveal details of how Muse Wonder's new models are made and trained.
However, Muse Wonder was built to compete with the best. Meta says that the largest version of the new model outperforms Anthropic's second-best model, the Claude Fable 5.1, and ranks below its best, the Claude Opus 5.5, in the AI index, a widely used score of the model's capabilities.
Meta says it outperforms OpenAI's best model, GPT-6 Astra, and Claude Fable 5.1 and Claude Opus 5.5 on some specialized benchmarks, including those that measure a model's ability to work with legal documents and build software. Benchmarks are not always a reliable measure of real-world performance, but this is still an important step forward for meta-models.
Muse Wonder also includes two new features designed to help make agents more effective and less prone to abuse. The new features, called active memory and active focus, are designed to keep the model locked on a set target and prevent it from becoming disoriented. For example, when AI is conducting complex research or building a large software project, this can be a challenge and can lead to costly mistakes or compound errors.
Meta says it has made a significant effort to train Muse Wonder to refuse to produce dangerous output, such as recipes for making bioweapons or hacking aids. The company says it scores well on several metrics designed to measure how resilient models are to these demands.
A series of recent incidents involving rogue AI agents from Anthropic, OpenAI, and others also show that powerful AI models can easily go astray when given open-ended commands. Meta did not reveal whether the agents employed by Muse Wonder misbehaved in testing, or what specifically it did to mitigate the problem.
In a post on X last month, Zuckerberg said the company had delayed the release of its Muse models to work on safety. Meta also introduced a framework designed to ensure the safe scaling of its models' capabilities in a blog post earlier this month.
However, Muse Wonder may add more fuel to a furious race to develop more capable artificial intelligence. It certainly shows that Meta is once again competing in the high-stakes AI race.