On October 12, Meta announced the Muse Wonder and Muse Wonder Full artificial intelligence models, describing the series as the largest and most powerful in its history. The models are intended to operate agents capable of carrying out multiple-step daily and professional tasks, rather than merely generating short text responses.
Meta has begun making Muse Wonder available to developers through the Meta Model API and the Muse Code programming platform, while Muse Wonder Full is available to Muse Code subscribers. The company also plans to integrate the model into its personal assistant, Muse, which it launched in the United States in September to perform tasks such as sending email and booking travel. The rollout will begin with service subscribers.
Memory Across Conversations and Continuing Work After Setbacks
The model focuses on two capabilities Meta calls Active Memory and Active Focus. The former enables it to retain important information across multiple conversations and recall details related to a task, including habits, preferences, and some content from previous conversations. The latter is intended to keep the model focused on its goal while carrying out complex work, trying alternative approaches, recovering from errors, and requesting user confirmation before important operations.
Meta says Muse Wonder was built on the Muse Spark foundation it announced in April, with expanded pretraining and reinforcement learning. The model was trained at the Prometheus data center, which has gigawatt-level capacity. The company says it demonstrated better-than-expected performance on unfamiliar tasks compared with the trends observed for Muse Spark 1.3.
From Contracts and Claims to Multi-Step Applications
Meta expanded the model’s knowledge in law, finance, and healthcare, and presents use cases such as processing insurance claims, reading leases, and managing a household budget. It also demonstrated examples involving planning a running program across a browser and multiple applications, designing products based on Instagram images and then ordering them, preparing a wedding seating chart according to complex conditions, and creating a playable 3D game from an image.
The models support text, image, video, and audio input, with text output and a context window of up to one million tokens. The Muse app will eventually use the model to create more complex outputs and develop specialized models available to all users, but Meta has not specified when it will be integrated or provided details about the plans that will include it.
Strong Results in Some Tests and Limits That Still Require Verification
According to Meta’s comparisons, Muse Wonder Full scored 13.1 on the Harvey LAB-AA test for legal tasks, compared with 8.6 for GPT-6 Astra and 4.2 for Claude Opus 5.5. It also scored 79.0 on DeepSWE 1.1, which measures long-horizon coding, ahead of Claude Opus 5.5 at 74.2 and GPT-6 Astra at 74.1.
However, the model did not lead in every test; Claude Opus 5.5 achieved the highest score on HealthBench Pro, Terminal-Bench 4.0-AA, ProgramBench, and the overall Artificial Analysis index. Some of the comparisons presented in the article also rely on evaluations published by Meta, while the full safety testing report remains pending publication.
Safety Before Expanding into Sensitive Tasks
Meta said that its assessment under the Meta Superintelligence Scaling Framework classified the risks of cybersecurity and loss of control, before mitigation measures, as medium or lower. In the chemical and biological domains, however, it did not rule out reaching the high-risk threshold. It therefore added safety training and system-level classifiers, and says the final assessment became medium or lower in all domains.
What changes in practice? Muse Wonder presents a vision of an agent that remembers context and continues carrying out a task across multiple tools and applications, which could reduce the need to divide work manually. However, its practical usefulness will depend on when it becomes available in Muse, the scope of supported plans, and the results of independent evaluations, particularly as the model moves into sensitive legal, financial, and healthcare fields.