Meta Launches Muse Wonder, a Frontier Model for Personal Agents

Summary

Meta executives Mark Zuckerberg and Alexandr Wang present Muse Wonder as the next, substantially larger model from Meta Superintelligence Labs, following the Muse Spark line. They say it is designed specifically for personal and professional agents rather than only for general chat, with active memory, stronger persistence on complicated tasks, multimodal perception, domain expertise, and frontier-level coding. The model’s active memory is intended to retain information mentioned over days, weeks, or months and use it in later tasks, while its “active focus” is described as the ability to complete long-horizon work, recover from setbacks, and find alternative solutions. Meta says Muse Wonder performs strongly on internal personal and professional agent evaluations and is state-of-the-art on DeepSWE and SWE-Marathon, while also performing well on Terminal-Bench 4 and domain evaluations in areas such as finance and law. The executives emphasize that internal evaluations and out-of-distribution performance matter more to product development than chasing saturated public benchmarks. Muse Wonder is expected to power difficult tasks in the Muse product, initially through paid subscriptions because it costs more to serve than Muse Spark. It is also planned for the Meta Model API, OpenRouter, and Muse Code; a higher-end Muse Wonder Full model will be available to Muse Code subscribers. Meta says the model was developed through a long pre-training and post-training process involving multimodal and domain data, supervised fine-tuning, reinforcement learning, carefully designed environments, and alignment work conducted throughout the RL process. The discussion also describes Meta’s broader infrastructure investment, including large compute clusters and tens of billions of dollars in frontier-model costs. Zuckerberg and Wang argue that AI products will remain differentiated because models can be optimized for distinct uses, and they expect personal agents to support tasks such as shopping, commerce, coding, voice interaction, and future hardware experiences.