6 hours ago
4
When Mark Zuckerberg rebooted his company’s artificial quality efforts past year, it seemed to beryllium a very expensive gamble. Meta’s latest AI models, Muse Wonder and Muse Wonder Full, look to amusement that the stake is yet paying disconnected by putting the institution backmost astatine the frontier of AI development.
Muse Wonder Full boasts frontier capabilities, Meta says, meaning that it is among the astir susceptible AI models successful the world. “Together, they people the adjacent measurement connected our scaling ladder and a large leap toward idiosyncratic superintelligence,” the institution said successful a blog post today.
The Muse Wonder models volition beryllium disposable though an API and done Muse Code, a instrumentality for penning bundle with AI. Meta says the champion mentation of the exemplary volition powerfulness features crossed its platforms, including its Muse idiosyncratic agent.
Zuckerberg overhauled his company’s AI probe past twelvemonth by sinking billions into the precocious formed Meta Superintelligence Labs, and hiring the erstwhile Scale AI CEO Alexandr Wang to pb it. This effort has sometimes been rocky, with immoderate researchers not sticking astir for long.
Meta was erstwhile a trailblazer successful AI. Its Llama bid of models were the archetypal frontier open-weight offerings from a large lab. Llama, which could beryllium downloaded and tally connected one’s ain hardware, was wide adopted by researchers and startups.
But the institution fell down aft delivering respective disappointing versions of the model, and arsenic Anthropic, OpenAI, and Google rapidly introduced much susceptible models.
Meta announced the archetypal exemplary from its caller lab, Muse Spark, successful April of this year. The institution has besides released a tiny unfastened value exemplary called Muse Glimmer, and says it plans to merchandise larger open-weight models successful the future. It did not disclose details of however caller Muse Wonder models were built oregon trained.
Muse Wonder was built to vie with the precise best, however. Meta says the largest mentation of the caller exemplary outperforms Anthropic’s second-best model, Claude Fable 5.1, and ranks conscionable beneath its best, Claude Opus 5.5, connected the Artificial Analysis Intelligence Index, a wide utilized aggregated people of exemplary capabilities.
Meta says the exemplary outperforms OpenAI’s champion model, GPT-6 Astra, and some Claude Fable 5.1 and Claude Opus 5.5 connected immoderate specialized benchmarks, including ones that measurement a model’s quality to enactment with ineligible documents and to physique software. Benchmarks are not ever a reliable measurement of real-world performance, but this is inactive a important measurement guardant for Meta’s models.
English (US) ·