Meta launches Muse Wonder models claiming frontier capabilities

Meta Platforms released two new artificial intelligence models on Monday called Muse Wonder and Muse Wonder Full. The company says these are its most capable versions for running agents that handle personal and professional tasks. This release marks the first time the company has announced a model it calls frontier technology.

The models include features called Active Memory and Active Focus to improve performance on long jobs like insurance claims. Active Memory keeps context from past conversations while Active Focus helps the system recover when it gets stuck. Meta states the software will check with users before completing important actions without warning.

Chief AI Officer Alexandr Wang told reporters the models match ChatGPT and Claude on coding and legal work. However, reports note no external benchmarks have been published to verify these claims yet. Meta’s internal testing compares them to GPT-6 Astra and Anthropic models but cautions that competitor runs were not optimized.

In specific tests like the Legal Agent Benchmark, Muse Wonder Full achieved passing results on 13.1 percent of tasks compared to 8.6 percent for a rival. Coding scores varied by test type with some showing Meta ahead while others showed competitors winning first place. The model card notes that some comparison figures are not controlled and may not reflect standard conditions.

A full safety report is promised soon but current warnings suggest the model could reach high-risk thresholds for chemical and biological threats without safeguards. Meta says it rates as moderate or lower on risk domains like cybersecurity when those protections are active. An agentic misalignment test found a zero percent rate of harmful actions to advance goals.

Developers can access the standard Muse Wonder model through the Meta Model API with published pricing tiers. The more powerful Muse Wonder Full version is restricted only to subscribers of the Muse Code service. This limits availability for users who do not pay for coding features.

The models were trained at the Prometheus data center in central Ohio which came online earlier this year. This launch follows a restructuring of Meta’s AI efforts led by Alexandr Wang last August. Mark Zuckerberg has committed hundreds of billions to infrastructure to support these systems.

It remains unclear if Meta will publish the model weights which determines obligations under the EU AI Act. The company previously declined a code of practice written for high-risk systems and killed another product after privacy issues. Shares rose as much as 4.7 percent following the announcement of the new technology.

Update (Help Net Security): Help Net Security reports the models can process images, video and audio in addition to text. Help Net Security notes Meta acknowledges specific weaknesses in spatial understanding and multi-step reasoning errors during evaluations covering finance and health tasks. Testing includes FocusBench for persistence and SWE-Marathon for long-running software projects, according to Help Net Security.