Meta has introduced Muse Wonder, a family of AI models it says can compete with the most capable systems available. The company describes Muse Wonder Full as a frontier model. “Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence,” Meta said in a blog post.
Meta says the largest Muse Wonder model ranks below Anthropic’s Claude Opus 5.5 and above Claude Fable 5.1 on the Artificial Analysis Intelligence Index, an aggregate measure of model capabilities. It also says Muse Wonder beats those two models and OpenAI’s GPT-6 Astra on some specialized tests, including tests involving legal documents and software development. Benchmark results do not always reflect how a model performs in real use.
Meta created a new AI lab last year
The release follows a major change to Meta’s AI research effort. Last year, Mark Zuckerberg put billions into the newly formed Meta Superintelligence Labs and hired former Scale AI CEO Alexandr Wang to lead it. Some researchers have since left after short stays. The lab’s first model, Muse Spark, arrived in April of this year.
Meta previously led a push toward open-weight AI with Llama, which researchers and startups could download and run on their own hardware. Later Llama releases disappointed, while Anthropic, OpenAI and Google introduced more capable models. Meta has also released a small open-weight model called Muse Glimmer.
Muse Wonder adds Active Memory and Active Focus, two features intended to keep an AI agent focused on its stated goal during complex work. Meta has not disclosed how the new models were built or trained. It says the strongest version will power features across its platforms, including its Muse personal agent.
Meta trained the model to refuse dangerous requests
Meta says it trained Muse Wonder to refuse dangerous requests, including instructions for building bioweapons or hacking. The company says it scores well on several tests of resistance to such prompts. Recent incidents involving agents from Anthropic, OpenAI and others have shown that AI systems can stray when given open-ended tasks. Meta did not say whether Muse Wonder agents misbehaved in testing or describe measures aimed specifically at that issue.
Zuckerberg said on X last month that Meta had delayed its Muse releases to work on safety. Earlier this month, the company published a framework for scaling model capabilities safely. Meta says it plans to release larger open-weight models in the future.
