Artificial Intelligence
Meta launches Muse Wonder AI models
Meta said it had released Muse Wonder and Muse Wonder Full, with access through the Meta Model API and Muse Code, and described the models as leading on agentic, memory and coding tasks.
- Meta introduced Muse Wonder and Muse Wonder Full as a new model series for personal and professional agentic tasks.
- The company said Muse Wonder was available through the Meta Model API and Muse Code, while Muse Wonder Full was available to Muse Code subscribers.
- Meta said the models scored strongly across benchmarks for agents, memory and coding, and that the release built on earlier work around Muse Spark.
- The company said the models were trained at its Prometheus data centre and would later power features in Muse.
- Meta said Muse Wonder had stronger safeguards against CBRN and cyber misuse, while keeping defensive cyber capability available under its scaling framework.
Meta introduced Muse Wonder and Muse Wonder Full as part of what it called its Muse Wonder series. It said both models were available that day, with Muse Wonder also set to appear in Muse and to support other future model work.
The company said the models were its most capable yet and claimed they performed at the top of its benchmark set for personal and professional agents, memory and coding. It also said the release extended work it had described earlier in Muse Spark.
Meta divided the model behaviour into areas it called Active Memory and Active Focus. It said the first was meant to preserve useful context across conversations, and the second was meant to keep working through obstacles, seek clarification when needed and check before major actions.
For coding, Meta said the models were trained on long-running tasks intended to help them keep progressing through testing and revision. It also pointed to benchmark names including Harvey's Legal Agent Benchmark, FocusBench and SWE-Marathon.
On safety, Meta said the models were trained with resistance to adversarial instructions and careful handling of private information. It said Muse Wonder had its strongest protections so far against CBRN and cyber misuse, while still allowing defensive cyber use under the Meta Superintelligence Scaling Framework.
Named in this story
Companies
- Meta
- said it had introduced the Muse Wonder models
Products and systems
- Muse Wonder
- the new model released for agentic tasks
- Muse Wonder Full
- the larger model in the new series
- Meta Model API
- one route through which Muse Wonder was made available
- Muse Code
- the product and subscription path for access
- Muse
- the product that would later use Muse Wonder
- Muse Spark
- the earlier model Meta said this release built on
- Harvey's Legal Agent Benchmark (LAB)
- one benchmark Meta cited for legal tasks
- FocusBench
- one benchmark Meta used to assess focus
- SWE-Marathon
- one benchmark Meta used for long-running software work
- Meta Superintelligence Scaling Framework
- the framework Meta cited in its safety claims
Places
- Prometheus data centre
- the site Meta said trained the models
How the source tells it
The piece read as a product announcement, with strong self-praise and future-facing claims about frontier capability and superintelligence.
- hype loaded superlatives and frontier framing around the launch
- boosterism repeated self-assessment of the models as the company’s strongest and most capable
- future promise forward-looking claims about later rollout and future model use