Meta describes Muse Wonder’s context and task-handling capabilities

Meta says the models carry important user context across conversations and retrieve details relevant to the task at hand.[4] The company also describes them as trying alternative approaches, working through blockers and adapting to new information.[4] When needed, they ask for clarification and check in before important actions.[4] These descriptions outline how Meta intends the models to handle work that unfolds over multiple steps, while leaving users with a role in resolving ambiguity and approving consequential actions.[4]

Meta says Muse Wonder is available through the Meta Model API and Muse Code, and that it will soon power experiences in the Muse assistant, with subscribers getting access first.[2] The developer portal lists pricing of $4 per million input tokens and $20 per million output tokens; it separately lists cached input at $0.20.[5] Muse Wonder Full is available only through Muse Code’s Power Usage and Maximum Usage subscription plans.[5] Meta says Maximum Usage has an 80-times-higher usage limit than Everyday Usage.[5]

Both models accept text, images, video and audio as input and generate text as output, according to Meta’s model description.[6] Meta lists a one-million-token context length at inference.[6] The company says its evaluations include tasks in law, finance and health, and that the models support coding projects from small applications to complex software.[4] Meta’s posted benchmark results show the models topping or matching recent offerings from OpenAI and Anthropic.[3] Its developer portal also says Muse Wonder Full outperformed other models evaluated on DeepSWE and SWE-Marathon in independent third-party runs.[5]

Meta says the models were trained to avoid sharing private information without permission and to refuse requests for help with harmful acts.[2] The company says developers and security teams can use models like Muse Wonder to find flaws in their own code and strengthen their systems.[2] Meta says it will release a full report with additional safety evaluations later.[2] That report may help teams assess how the stated safeguards perform beyond the descriptions currently provided. Meta also says it remains committed to open source, but the evidence here does not specify what licensing or release terms will apply to these models.[3]