Meta releases its first frontier mannequin, Muse Wonder

Meta’s Superintelligence Lab launched a pair of new state-of-the-art models on Monday, Muse Wonder and Muse Wonder Full, representing the corporate’s first try at a frontier mannequin.

Where Muse Spark aimed to be a light-weight mannequin that would carry out at low latency, the Wonder fashions are bigger and mentioned to be extra succesful, with larger benchmark scores and higher outcomes on complicated duties like coding.

According to Meta’s testing (offered in the model card), Wonder and Wonder Full obtain comparable benchmark scores to Fable 5.1 and GPT-6 Astra.

“Together, they mark the following step on our scaling ladder and a significant leap towards private superintelligence,” Meta’s weblog publish concerning the fashions reads.

That scale is mirrored within the pricing: Accessed by the Meta Model API, Wonder will price $4.00 per million enter tokens and $20.00 per million output tokens, making it roughly equal in price to GPT-6 Sol or Opus 5.5.

Notably, Wonder Full will solely be accessible to Muse Code subscribers, and won’t be served by the API. Both fashions may also be used to energy the Muse agent.

In holding with Meta’s curiosity in growing private assistants that may carry out duties over an extended interval, the corporate positioned specific emphasis on the brand new fashions’ capability to retain and synthesize data given by the consumer, a property Meta calls “energetic reminiscence.”

To obtain higher recall, Meta highlighted the MemoryBench benchmark, which challenges continuous studying and reminiscence development frameworks. According to the corporate’s testing, Muse Wonder now scores larger on MemoryBench than Fable, Opus or Astra.

“We put an incredible quantity of effort into ensuring your mannequin can bear in mind what’s essential to you over days, weeks, months and past,” mentioned analysis scientist Kavya Kopparapu in a video accompanying the launch.

The new fashions are the primary frontier mannequin launch since Scale AI founder Alexandr Wang restructured Meta’s AI efforts in August of last year. Since then, Meta has launched its light-weight Muse Spark mannequin and Muse agent, however has prevented the type of state-of-the-art mannequin releases OpenAI and Anthropic are identified for.

According to the publish, the fashions had been educated at Meta’s Prometheus data center in central Ohio, which got here on-line earlier this 12 months.