Meta 发布 Muse Wonder 与 Muse Wonder Full 智能体模型

Meta has launched Muse Wonder and Muse Wonder Full, two AI models built for personal and professional agents, with capabilities spanning long-term memory, multi-step task execution, and coding.

The models are part of Meta’s effort to scale its AI efforts. Meta trained both models were trained at the company’s gigawatt-scale Prometheus data centre.

The company also plans to use the model to develop specialised models for wider access through Muse.

The Muse Wonder series combines two capabilities that Meta calls Active Memory and Active Focus. Active Memory allows the models to retain relevant information across conversations and retrieve it when needed for a task.

“One of the most critical steps towards useful personal and professional agents is reliable memory,” Meta said in a release. The company describes the capability as “carrying important context about you across conversations and actively retrieving the details relevant to the task at hand”.

Active Focus refers to the models’ ability to work towards a goal over an extended period, try alternative approaches, address obstacles, and adapt to new information. Meta said the models can ask for clarification when required and check with users before taking important actions.

These capabilities aim to support agents that can work across applications and handle tasks involving multiple requirements and constraints.

The company said the models achieved state-of-the-art results across benchmarks covering personal and professional agents, memory, and coding. Its evaluations include AutomationBench-AA for business workflows, MemoryBench for long-term contextual memory, SWE-Marathon for sustained software engineering, and DeepSWE for long-horizon coding.

Meta also evaluated the models on specialised tasks in law, finance and health. These include Harvey’s Legal Agent Benchmark, which tests whether models can work with source documents and follow detailed instructions to produce legal work, and Vals Finance Agent V2.

Meta said Muse Wonder Full can generalise more effectively to unfamiliar tasks than its earlier Muse Spark 1.3 model, based on the company’s evaluation of performance as reasoning effort increases.

The company examined the relationship between performance on tasks targeted during post-training and performance on tasks that were not targeted during that process. It found that Muse Wonder Full performed better on unfamiliar tasks compared with Muse Spark 1.3.

Meta attributed this to the stronger base models underpinning the Muse Wonder series, saying that improvements from post-training carry over more effectively to tasks outside the training focus.

“This versatility especially matters for personal agents, which need to handle many combinations of goals, tools, and constraints—not all of which we can anticipate during training,” the company said.

Meta said it trained the models to handle private information carefully and resist adversarial instructions, given that personal agents may access messages, calendars, and private documents while acting on a user’s behalf.

The company also said Muse Wonder has its strongest safeguards to date against chemical, biological, radiological, and nuclear risks, as well as cyber misuse. It said the models retain access to advanced cyber capabilities for defensive use under Meta’s Superintelligence Scaling Framework.